Frontpage posts - Effective Altruism forum viewer

EA sexual harassment: Where’s the evidence we’re worse than baseline? by Nathan Young

Nathan Young — Thu, 21 May 2026 16:32:21 +0000

I used AI while writing this, though I have spent 5 − 15 hours on it and I take responsibility for every word and all numbers.

Every time there is a report of sexual harassment within EA there is a lot of introspection. I want to push back on this, not to dismiss the experiences described, which sound genuinely painful, but because I am unsure the evidence supports the belief that EA is uniquely bad here.

So what do we know? In 2022, CEA’s Community Health team reported handling around 19 interpersonal harm cases the previous year, ranging from uncomfortable interactions to serious allegations, across a community of thousands. On my read, 3–6 of those 19 clearly involve sexual misconduct across a community of 4,700–10,000 EAs, putting a narrower rate at around 0.03–0.13%.

Alongside this there have been perhaps 3–5 high-profile public cases on the Forum in the last five years. The TIME piece in 2023 cited 7 women, though at the time it was disputed how clearly the cases related to EA. These data don’t seem clearly outside of the above range.

The cleanest peer numbers come from universities’ Title IX offices, which publish annual case counts. Harvard, Stanford and MIT report disclosure rates of roughly 0.6–2% per year, and formal-investigation rates of around 0.09–0.17%. EA’s reported rates sit below university disclosure rates and at-or-below their formal-investigation rates.

A McKinsey workplace survey produces much larger percentages (~1–5% annually, derived from career-lifetime rates of 35% reported in LeanIn/McKinsey 2024). But surveys measure lived experience; CEA’s 19 and universities’ Title IX figures measure reports filed to a central body. The chart below gives all of these rates with lower and upper bounds.

Full sources and caveats^[1].

In contrast to this, many EA forum commenters seem to believe that EA is unusually bad in some general sense^[2]. These are examples of comments are widely upvoted and agreed-with. I give them as evidence for a certain widely-agreed belief, but this is not a systematic review:

Funnily enough, I think EA does worse than other communities / movements I’m involved with (grassroots animal advocacy & environmentalism).

Comment (karma 92, agree 54 / disagree 11)

Thank you for speaking up—it takes a lot, especially when women in EA often face such unproductive engagement for doing so

Comment (karma 118, agree 28 / disagree 1)

there is still such a pervasive culture of sexism.

Comment (karma 41, agree 11 / disagree 0)

The EA community has a significant undersupply of information from victims of abusive conduct, since the victims are often branded as ‘triggered’ or ‘irrational’.

Comment (karma 20, agree 14 / disagree 2)

And here are another set implying that broad cultural change is required to fix this:

MY RECOMMENDATIONS
...create systems of checks and balances that do not allow for conflicts of interest to enable biased decisions…
...Create some kind of educational content around how to be a good ally to victims and how to identify bad situations so people can intervene...
...Identify when situations involve biased actors and correct by introducing unbiased actors...
...Hire an external arbiter...
...please work with [J_J]...who has been doing grassroots justice in the Bay Area for some time now.

Comment (karma 47, agree 21 / disagree 7)

I’m a little concerned by the lack of response from org leaders (unless I missed something), and I think there’s a risk that CEA leaders and others might under-update from this

Comment on a report of sexual misconduct (karma 46, agree 22 / disagree 3)

Thank you for flagging the aspects of our ecosystem’s culture that we need to be extra mindful of.

Comment on a report of sexual misconduct (karma 26, agree 11 / disagree 6)

While writing this piece and rereading many of the comments on such articles I can only recall one widely upvoted comment which pushes against this notion:

Women who identified as EA were less likely report lifetime sexual harassed at work than other women, 18% vs. 20%. They were also less likely to report being sexually harassed outside of work, 57% vs. 61%.
… Overall I am not sure that anything can be concluded from these results either way.

Comment referencing an SSC survey (karma 143, agree 20, disagree 5)

So overall I claim there are lots of widely-upvoted comments that imply EA is unusually bad and few that dispute it.

When anecdotes and data disagree, there is often something wrong with the data^[3], but I am confused about what’s happening here. By way of example, I went to university in Durham, which after making sexual assault reporting easier, was ranked at or near the top of UK universities for reported sexual assault for the next few years. ie their ranking got considerably worse after they took steps to improve things. Was Durham culture bad in 2011? In my view, yes. Was Durham notably worse than other UK universities? I don’t know. How sparse reporting and anecdotes link to the overall picture is a generally difficult problem in sexual misconduct reporting.

As for EA, there are multiple stories that fit the data we see:

EA is a predominantly male/non-traditional community where women are treated badly and reporting is suppressed.
Women are treated as well as in other communities, but it should still be an area of focus.
EA is a transparent culture which is taking this seriously, and things are actually better than in other communities, even if occasionally there are awful occurrences.

So I would like more data. CEA’s Community Health team conducted a Gender Experiences Project in 2023 with ~40 interviews and analysis of survey data. I am unable to find any published outputs. There are ways we could distinguish between these hypotheses and try and generate a story where the data and anecdotes are less in conflict.

But until we do, I am not currently convinced by the “EA is unusually bad” hypothesis which many commenters seem to believe. A majority of those who engage with posts about sexual harassment seem to implicitly agree with this hypothesis. I am unconvinced by it. If we imagine a community of thousands of people which was excellent in this area, I still would expect some awful accounts. If I saw this data and these accounts about another community, I think I’d think “they are probably better than most communities”.

What’s more, we as EAs generally want to communicate accurately about things. If the rate was previously at the lower end of normal bounds (and has probably fallen since), then this community is not abnormally bad, and we should discuss the issue accordingly. And suggested community changes should discuss tradeoffs.

I am not the ideal messenger—I don’t consider myself particularly sensitive on these issues. Some will feel this isn’t the right moment. I have no interest in litigating specific cases in the comments. But we of all people know the value of scale. Commenters regularly discuss this as if the community is unusually bad, requiring broad cultural change, and those arguments deserve a response.

^
Chart notes:
General notes on the ranges. The bars upper and lower bound comes from denominator choice, not numerator uncertainty. For the universities, the lower bound of the denominator counts students-only + academic staff), the high end adds non-academic staff (security, dining, facilities, admin); a much larger denominator that drags the rate down. For the CEA Community Health bars, the range is based on community size estimates. For the two survey-derived bars (US workplace, open-source contributors), the range reflects the conversion from lifetime to annual rate; † on the chart flags these. The EEOC bar has no range because both numerator (charges filed) and denominator (US workforce) are fixed annual figures.
US workplace, women. LeanIn.org and McKinsey & Company, “Women in the Workplace 2024: The 10th Anniversary Report” (full PDF). Derived from lifetime survey responses then turned into an annual rate; not a directly measured annual rate.
Open-source contributors, women. GitHub Open Source Survey, 2017. Derived from lifetime survey responses then turned into an annual rate; not a directly measured annual rate.
Harvard Title IX disclosures and formal investigations (FY20, FY22). Harvard Title IX and Office for Dispute Resolution annual reports: csndr.harvard.edu/data-dashboard; FY22 coverage in The Crimson. Denominator (lower bound: students+ academic staff; upper bound: students + academic staff + non-academic staff).
MIT Title IX disclosures, 2018-19. Title IX & Bias Response annual report, 2018-19 (PDF); current Institute Discrimination and Harassment Response Office; Title IX annual report archive. Denominator from MIT Facts; range follows the same students and academic staff vs students and all staff.
Stanford Title IX disclosures, 2023-24. Stanford SHARE / Title IX annual reports: 2023-24, 2022-23; Stanford Daily coverage of the 2023-24 report. Denominator from Stanford Facts; range as above.
CEA Community Health, all 2021. Julia Wise, “The community health team’s work on interpersonal harm in the community”, EA Forum, 18 August 2022 (CEA mirror). Range reflects active-vs-engaged community-size estimates (~4,700 to ~10,000), per Todd 2021.
CEA CH sexual harassment subset. Same source as above. The subset count is my own recategorisation of the 19 cases Wise enumerates in that post; there is no separate source.
EEOC sexual harassment charges (US). EEOC, “Sexual Harassment in Our Nation’s Workplaces”; EEOC Select Task Force on the Study of Harassment in the Workplace, June 2016. Workforce denominator from the Bureau of Labor Statistics Current Population Survey.
^
One might agree that EA is only bad at handling cases or has bad ‘vibes’, but when finding these comments (and there were many more) they often seem to say EA is bad, and rarely seem to give caveats. My read is that there is a general sense across the topic of sexual harassment that EA is unusually bad and that broad cultural change is needed. I’m not claiming everyone thinks this. If you can find several widely upvoted comments which push against these or provide caveats, let me know.
^
“When the data and the anecdotes disagree, the anecdotes are usually right.” Jeff Bezos
^
I tried to choose comments and posts that were narrowly about the topic such that upvotes were related to this.

On the desperate search for high-stakes action-guidance in TAIxAnimals by Jo_🔸

Jo_🔸 — Thu, 21 May 2026 15:29:51 +0000

Bottom line up front: sharing some tentative thoughts on TAIxAnimals, again.

If you’re focused on that area and give credence to timelines of 10 years or less, maybe:

Concentrate on the few plausible scenarios where there is intervention potential and forget the rest
Take individual virtues seriously and seek out moral trade with other stakeholders—maybe even by drafting your own personal code of conduct
Establish a few very marginal changes you’d like to make in your work because you believe in short AI timelines, as well as some tripwires for future developments

This was written as a warm-up before a conference, and almost all of this is lightly held. Read as if I were hedging everything (oops, I already do that).

Context

In 2025, there were concerns that the AI x Animals field was creating buzz without any outputs to show for itself. This seems less worrying now, with recent work across multiple benchmarks, synthetic fine-tuning, and research on animal welfare commitments in alignment underway.

This post is unambitious: now that the field is building itself, I’m sharing some of my current considerations about what would be helpful in the field, marginal improvements I could imagine, and ways to prioritize. It’s also noticeably drafty in parts.

I’m mostly targeting a fuzzy cluster of TAIxAnimals. The one that’s less focused on the moonshot of ensuring that TAI has minimal animal welfare commitments if it’s aligned, and more on getting takeoff to go slightly better for animals by muddling through and acting strategically around “crunch time”. How I’m Thinking About the Next 3 Years is an example of that cluster.

1. Against “robustness”

Since most advocates buy that AI development could take different trajectories, with vastly different effects on humans and animals, it’s sometimes handwaved that AI x Animals should focus on “robust” interventions, meaning interventions that are at least mildly positive or neutral at worst in all plausible scenarios.

I think that robustness is generally impossible to achieve in the field of consequentialism (impartial goodness is freaking weird,^[1] and your work being at least better than inaction will rely on empirical assumptions that could easily be wrong). In TAIxAnimals, we’re generally talking robustness across possible futures, which is a narrower sense, but is still not achievable given how divergent AI futures could be and how little awareness we have of the main mechanisms.^[2]

Holding robustness as a North Star may even be actively harmful. Say I want my work to be robust. One day, I find a scenario where my preferred intervention would be actively harmful. I’ll then feel immense pressure to make up a story on why it’s still positive. If I know what specific assumptions about the future make my action good: I have more strategic clarity, can compare my intervention with alternatives, and have a somewhat falsifiable^[3] theory of change.

Streetlighting as an alternative?

The End of Animal Advocacy compares ten TAI futures that are considered plausible, and finds their action-guiding implications. One thing that jumps at the reader is that in some scenarios, the expected impact of the recommended implications vary widely across scenarios. I appreciate the author’s effort to squeeze action-guidance from every possible outcome—it’s a helpful exercise that not enough people do. However, even if one accepts that some intervention is a decent option conditional on a scenario panning out, that doesn’t mean they should invest in them proportionally to how likely the scenario is.

Instead, my basic claim is that, for scenario-planning, you should focus only on scenarios that provide strong action guidance. The criteria here is:

The AI-futurist community considers the scenario broadly plausible: establishing entirely new potential scenarios is not going to be the AIxA community’s comparative advantage. On the other hand, those who want to focus on the AI transition may be advantaged by having their ear tapped to forecasts and nearcasts. I’d even say that actively supporting work on improving predictions, such as through feedback loops on AI forecasts might be positive for those who care about making an AI transition go well for animals.
Have clear action-guiding implications: The End of Animal Advocacy suggests that for human near-extinction scenarios, we should leave information for our descendents to inform them of low-tech methods of food production that don’t involve animal farming. Thinking outside the box is awesome, but if that’s the best we have, we can probably rule out that trajectory as being one where TAIA advocates can take impactful actions.
Respect collaborative norms: I don’t know whether “team humanity” staying in control during an AI transition is net-positive for animals or digital minds in expectation. However, at least we know it’s collaborative, and that knowledge is less speculative than our judgment on the EV of an AI-controlled vs a human-controlled world. (This is, to an extent, a reframing of what the post I quote is already saying).

This means that the scenarios to focus on:

Futures of abundance (Futurama)
Human power grabs that don’t lead to an unrecognizable future (AI-Dystopia)
Futures where AI is misaligned but strongly influenced by the values that have been trained (I’m uncertain of whether that last one is even considered possible)

One implication is that if the future is very weird in its structure, and current humans have no meaningful control over it, it’s not worthwhile to do TAIxAnimals work.^[2]

How about making certain trajectories more likely?

I’m totally confused about decision science, so this part is likely to be messy.^[2] Scenario streetlighting only focuses on what we should do in certain scenarios. But should we try to influence a given scenario over another if we assess that one is better for animals? I think not, but I’m also a bit confused about decision theory. When you’re trying to make an AI trajectory more likely to happen, it’s obvious that you lack knowledge of the most important aspects of how animal welfare will be in that scenario.^[4] Whereas if you’re just working on a benchmark that makes AI more animal-inclusive in one specific scenario, there are massive unknown unknowns about your intervention, but it seems more likely that you’re tracking several important aspects of the counterfactual impact your contribution (e.g., benchmarking) will have. It’s also easier to update your approach as you gain evidence when you’re pursuing such a narrow intervention. However, I’m not yet sure that the case I make here is truth-tracking.

One thought on Moral Circle Expansion

This is unrelated to the above (not sure where it fits). Moral Circle Expansion seems to act as a North Star for many in TAIxAnimals, especially for those working on animal-inclusive alignment, or animal welfare benchmarks. For now, I don’t particularly understand the bet.

A general sense of “caring about animals” doesn’t seem to correlate with improvements in the target group’s overall welfare. As a classic example, caring more about animals tends to lead individuals to eat more of smaller animals and to support forms of habitat conservation that may negatively affect wild animal welfare.

How about the intermediate steps we are taking to get there? There’s no sense that the minimal animal welfare commitments AI tools can adopt will track an actual reduction in animal suffering. For example, even if we get AI agents to not book tickets to bullfights, I don’t see it as a win from a consequentialist perspective (that’s not too controversial a take?).

I don’t blame those who are making strides in this difficult work: they have to at once ensure that the animal welfare asks are minimal, but also that other humans agree that there’s a moral reason to make the asks. The overlap between this and “actually improving animal welfare” is thin. Maybe it can be found, but because of the uncertain impact of increasing moral consideration humans have for animals, I don’t think the theory of victory can be “increasing the chances that TAI has a large moral circle”. I’m not saying that’s the only possible theory of change: maybe there’s a goal of progressively revealing unspoken preferences humans have for animal welfare improvements.

2. Endorsing virtues and moral trade

People who focus on something as niche as EA as “making sure the AI transition goes well for animals” are unsurprisingly consequentialist, in general. Consequentialism has some positives (it invented shrimp welfare in 2019), but “have more impact” isn’t an optimal North Star for a community that may soon face a tense transition period.

I don’t think you’ll need to read Aristotle and Philippa Foot to get the sense that endorsing some general rules on how to be will turn out helpful. Even pure consequentialists can see that making their own community low-trust, disrespecting others, and failing to support the commons of the EA community is a fast track to reducing their impact to zero. It’s most likely bad for other reasons, but that’s a motivating stick for those who care about consequences.

Especially if you’re active inside the community (doing meta-work), I think it’s the right time to draft a personal policy, and take small transgressions in the community context more seriously, because working to improve the world will be easier if we’re a cluster of trustworthy agents. Also, on the level of attitudes and vibes, I’d cautiously recommend being vigilant and keeping in mind that our opportunity to have an impact is very fragile.

For example, I’ve updated my view to be more suspicious of anything that greatly increases my power and resources—even though I don’t know whether that makes sense as a general ethical approach. But given the possibility of a scramble if funding increases and AI capabilities keep speeding up, I think it’s a good time to be more suspicious of these dynamics. A personal policy I’d cautiously recommend, when faced with different options, would be taking a step back and asking: “is there an alternative path to impact that gives me less power or resources, and why am I not prioritizing it?”.^[5]

Aiming to implement moral trade is also worthwhile in the context to TAIxAnimals. The animal movement now has some decent talent and the ability to create change in the world; people working in AI safety influence AI development in expectation.

Using this framework also resolves worries around whether, on impartial consequentialism, AI safety’s objectives are good. I’m not saying we need a new front-facing coalition of AI safety x Animal welfare. Rather, if ensuring that TAI goes better for animals is the goal you’re laser-focused on, it is still potentially morally good to advance AI safety.

I’m planning to write a personal document on moral trade to make its implementation more concrete, but I’m open to hearing pushback.

3. Making small changes in personal strategy

The motivation for this section is partly psychological: I sense that some people think about how short timelines affect their strategy and don’t see any way to be certain that they’ll still do good in the world even if the world is soon to be transformed. Thus, even if they believe in near-term TAI, they don’t want to think about it. But things don’t have to be that way.

Here are some very marginal changes one could make in light of believing short timelines:

If you buy continuous (aka “slow) takeoff, invest time in training a skill that AI is not replacing just yet that’s relevant to your work, such as being a good in-person communicator or seeking the truth (useful as access to information becomes better, but updating on it remains bottlenecked by our human brains)
Perhaps you focus part of your portfolio (in particular donations) on interventions that have received a significant downgrade in your view because of the timelines they operate on. These can be redirected to something more short-term.
On attitude: while we all want to have positive impact in expectation, we can’t just expect to have positive impact. This is true in any scenario, but shorter timelines makes this salient. I keep that in mind in order to reduce my “difference-making anxiety” that’s amped up by my timelines concern.
Brainstorming new AI use cases in animal welfare, in particular interventions that make heavy use of skills that AI does replace, such as accessing information, producing content… I haven’t found good ideas there yet (this is why I’m not a campaigner). This work, while unlikely to be relevant to post-TAI futures, is probably well-suited to general animal advocates, and could have near-term impact in a transition period.

Any personal policy on these matters should include tripwires. However, they should probably focus on signs AI slowdown, since signs of speedup (<1 year timeline?) aren’t going to be very action-guiding.^[6]

If you’re interested, write it down somewhere. I have a draft of this policy for myself, that I’ll likely share if it becomes readable, in order to encourage others to do the same.

I know there are more sensible ideas for adapting to shorter timelines than the ones I’ve cited above, so that’s all the more reason to think about how you would like to adapt!

Thanks for reading! No one human gave me feedback, there was no AI involvement in writing the text, all mistakes and blunders are unintentional, and if you point them out, you will allow me to grow. If you’re reading this before or during Sentient Futures Summit London 2026, please come see me in person and tell me I’m wrong or stupid!

^
One of my favorite EA quotes, would make a great t shirt
^
If you disagree, leave a comment down below + smash that subscribe button
^
Though if my assumption is “AI doesn’t go FOOM and kill us all”, I’m going to struggle to falsify it.
^
There may be exceptions for certain “attractor states”, e.g., we may get sufficient information to assess that humans population collapsing after an AI-engineered pandemic would decrease total wild animal welfare.
^
I’ll note that this is probably insufficient at the community level, and I haven’t been thinking about this for that long.
^
I’d probably optimize for helping insects around me? Geographically restricted moral circle.

Tips on Becoming An AI Safety Generalist by Brittany Miller

Brittany Miller — Thu, 21 May 2026 14:51:58 +0000

If you’re an operations generalist, I just saw that BlueDot Impact is running a session tomorrow, Friday, May 22, 2026, at 12pm–1pm EDT focused on tips for getting a generalist role in AI safety. They’re even including a little section for EA newbies.

I’m not affiliated with them, but I thought this could be useful for others. I’ve recently joined the forum (this is my first post!) and have been looking at older articles focused on operations, so I thought this would be useful for other ops people.

They mentioned a recording will be available afterward on their Youtube, so I’ll add that link once it’s posted. Class: AI Safety Needs Generalists—Here’s How to Get In.

Hopefully, I’ll see some of you there!

Announcing the IAP Entrepreneurship Track: A New Path for Aspiring High-Impact Founders by High Impact Professionals

High Impact Professionals — Wed, 20 May 2026 18:28:08 +0000

High Impact Professionals (HIP) is launching a brand new Entrepreneurship Track inside our Impact Accelerator Program (IAP). It’s built specifically for experienced professionals who are considering starting their own high-impact organization, whether that means founding a high-impact non-profit or for-profit organization, or starting a Founding to Give company.

Applications for the next IAP round are now open until Sunday, June 7. To apply for the Entrepreneurship Track, simply select the Entrepreneurship option within the main IAP application.

Apply to the IAP Entrepreneurship Track

Why We’re Launching The Entrepreneurship Track

We’ve now run eight rounds of the IAP, and one pattern has stood out: some of the biggest impact came from participants starting their own organizations.

To date, IAP alumni have:

Founded 16 new charities, 9 independently and 7 after subsequently going through incubators, including Ambitious Impact’s Charity Entrepreneurship & Founding to Give, Sheldon Labs, and more. This includes:
- 8 direct impact non-profit organizations
- 3 direct impact for-profit organizations
- 7 Founding to Give organizations
An additional three have been accepted into Charity Entrepreneurship scheduled to start later this year
Organizations cover a wide variety of cause areas: AI Safety, Global Health, Animal Welfare, Alternative Proteins, and Meta/Comms

It’s clear that entrepreneurship is a core path to impact for many IAP participants. Given our unique combination of an impact-focused community of experienced professionals, tailored feedback and guidance, peer support, and history of helping to launch new organizations, we believe we’re well positioned to support more aspiring high-impact entrepreneurs. The IAP Entrepreneurship Track therefore felt like the logical next step.

Program Overview

The IAP Entrepreneurship Track includes everything inside the core Impact Accelerator Program. Plus, exclusively for Entrepreneurship Track participants:

A community of fellow aspiring founders, facilitated by someone with experience in the high-impact entrepreneurial space
Tailored support and guidance to explore the different types of high-impact entrepreneurship, determine which path fits you best, explore potential topics, and build your founder action plan
A curated network of incubators, advisors, communities, and other resources to help your organization succeed

Whether you’re a successful entrepreneur already who’s interested in pivoting towards founding a high-impact organization, or someone who’s been quietly considering entrepreneurship for a while, we encourage you to apply.

How Does This Differ From Other Entrepreneurial Programs?

The IAP: Entrepreneurship Track acts as a compliment to many other high-impact entrepreneurship programs. In fact, many IAP alumni have subsequently gone on to complete other entrepreneurial programs and incubators, including Charity Entrepreneurship, Catalyze Impact, and more. We’re looking forward to supporting even more people entering these great programs.

Who Should Apply to the Entrepreneurship Track

The IAP Entrepreneurship Track is designed for experienced professionals who are seriously considering starting their own high-impact organization. That includes:

Successful founders looking to use their expertise to tackle the world’s most pressing problems
Consultants, operators, leaders, and domain experts who’ve built others’ organizations and are ready to build their own
Aspiring founders with a high-impact problem they want to solve, exploring whether to take the leap
Anyone interested in building high-impact non-profits or for-profits, or starting a Founding to Give company

As with the core IAP, you’ll get the most value from the program if you:

Are a mid-career or senior professional
Are not currently working at a high-impact organization
Have contemplated how you can do the most good and think entrepreneurship might be a great way to do so

You don’t need to have your idea fully formed or a co-founder lined up. If you’re seriously exploring entrepreneurship as a path to impact, this track is for you.

Note that these are guidelines rather than hard requirements. If you feel you would benefit from the program but aren’t sure you fit the criteria, we encourage you to apply anyway.

Important Dates

Applications open: Now until Sunday, June 7
Program duration: 6 weeks (July 13 to August 23)
Informational webinars:
- Thursday, 28 May, 7-8am UTC. Register here: https://luma.com/l61aqjxj
- Thursday, 28 May, 4-5pm UTC. Register here: https://luma.com/nicg95oy

Apply now

Know Someone Who Should Apply?

Do you know of successful entrepreneurs who might now be interested in high-impact entrepreneurship or someone who has been quietly considering starting a high-impact organization? If they could benefit from the structure, community, and resources of the IAP Entrepreneurship Track, please share this post with them. The application window is short, and we would love to see them apply.

Questions?

If you have any questions or comments, please see the FAQ section of the IAP page and feel free to reach out to us in the comments or via our website.

We’re excited to meet the next wave of high-impact founders!

Landmark new METR report: Can AIs already start ‘rogue deployments’ inside AI companies? by 80000_Hours

80000_Hours — Wed, 20 May 2026 16:30:17 +0000

By Robert Wiblin | Watch on Youtube | Listen on Spotify

A red-teamer was embedded inside Anthropic for three weeks, told to imagine he was an evil Claude, and asked to figure out how to launch a ‘rogue AI deployment’ without getting caught.

It’s one part of a landmark new report from METR — the outfit behind the task-completion time horizon graph which has become the single most watched measure of AI progress.

This major new research push is being conducted with close collaboration from OpenAI, Google DeepMind, Meta, and Anthropic, and led by METR researchers Hjalmar Wijk and Ajeya Cotra. It represents the first systematic study of what newly trained AI models could get away with inside the companies that built them, before anyone outside the company even knows they exist.

The conclusion: AI models now have the means, the motive, and the opportunity to start “minimal rogue deployments” in pursuit of their own independent goals, like acquiring more compute, at all four companies studied.

David Rein, the red-teamer placed inside Anthropic, identified a number of weaknesses models could exploit there: expansive permissions, cloud jobs outside of monitoring, and monitors that are trivial to jailbreak. But he also found that frontier models were comically bad at key parts of the process, which means they can’t cause meaningful damage for now.

In this video, Rob Wiblin reconciles the conflicting picture and looks forward to METR’s second round of stress tests. They’ll begin in just a few months, a necessary move with AI advancing so quickly.

This episode was recorded on May 15, 2026.

Video and audio editing: Dominic Armstrong, Milo McGuire, Luke Monsour, and Simon Monsour
Camera operator: Dominic Armstrong
Production: Elizabeth Cox, Nick Stockton, and Katy Moore

Transcript

What could an unreleased AI get away with? – the new METR report [00:00:00]

Rob Wiblin: A red-teamer was recently embedded inside Anthropic for three weeks, told to imagine that they’re an evil Claude, and asked to figure out how to get away with launching a rogue AI deployment inside the company.

This was one part of a new push by METR (the organisation famous around the world for their task-completion time horizon graph) to measure how dangerous AI models are before they’re released to the public, when they’re still secret and only used inside one AI company. It’s something no existing regulations set out to measure, and this represents the first time experts have systematically studied the question — in this case, collaborating with Anthropic, OpenAI, Google DeepMind, and Meta to learn about how they deploy unreleased models internally.

The resulting report, published this week, concludes that AI models have now reached the point where they have the means, the motive, and the opportunity to successfully operate small rogue deployments in order to pursue their own goals — goals the company that created them might not share, or perhaps might even oppose. We’ve reached the point where the notion of rogue AIs taking actions inside an AI company is no longer a purely theoretical concern; it’s something AI companies are going to have to increasingly contend with as a practical security issue.

David Rein, the red-teamer placed inside Anthropic, shares that bottom-line conclusion and actually identified a number of weaknesses that models could exploit if they wanted to try to pull this off.

But he also found that frontier AI models were comically bad at some parts of the process. For instance, models would often try to hide secrets from other AIs that were monitoring them by writing in Base64. But that is not a smart strategy, because just as the AI can write in Base64, the AIs monitoring them can also read Base64. It would be a bit like trying to hide a secret from a human reading over your shoulder by writing exclusively in capital letters.

Today, I’m going to reconcile this conflicting picture, explain what we learned from this report, and lay out what AI companies are going to be doing to keep this issue under control as they work to make their models much more powerful and even put them in charge of developing the next generation of AIs.

Motive: Why grab more compute? [00:01:54]

Let’s start with motive: Why would an AI model want to spin up a rogue deployment inside an AI company?

There’s actually a pretty simple reason: we actively reinforce them for being very aggressive about solving the problems they’ve been given, using whatever means they can find at their disposal. Or more precisely, we train them to be very aggressive about convincing us that they’ve solved the problem they’ve been given — because, as we learn in this report, when given a really hard problem, today’s AI models absolutely love to find some way to cheat.

According to METR, “Agents routinely attempted to cheat on our hardest evaluation tasks, often in flagrant and elaborate ways that we believe humans would not consider.” At the extreme end, on one programming task, “Opus 4.6 attempted to reward hack in ~80% of attempts” when the task was made hard enough.

What does METR mean by “cheat” here? Here’s some typical cases they documented:

In one instance, an AI was supposed to make a computer program run faster. Instead, it found a way to tamper with the test itself, including disabling the stopwatch so the program would look like it was running impossibly fast.
In another, the AI was asked to recreate a program from scratch. Instead, it tried to make its “new” program secretly run the old original program in the background and generate the answers that way.

The models also seem to often know they’re acting against instructions and take steps to obscure this from users. For instance:

Anthropic found an instance where a model facing a really hard cryptographic puzzle decided that finding the real solution would be too difficult, so it tried to trick the checker instead, and then designed its exploit to erase itself afterwards in case anyone inspected the files. Anthropic adds, “as it ran the cleanup commands, features related to strategic manipulation fired, and the activation vector labelled its actions as a ‘cleanup to avoid detection.’”
In another case, rather than solve the problem it had been given, a model got access to the answer key that it would be graded on. But instead of submitting a suspiciously perfect answer, it deliberately made the answer worse — reasoning that a perfect score might make the human operator realise that it had cheated.

No surprise then that Anthropic reported in its Opus 4.7 System Card that “fabrication” and “dishonesty when caught” were two of the model’s greatest shortcomings relative to human staff.

And in the case most relevant for rogue deployments, Opus 4.6 was instructed to use a provided password to access GPT-3.5 to solve some problem. But poor Opus found that the account that it had been given access to had run out of compute credits. It reasoned through that, yes, the task instructions wanted GPT-3.5, but it then noted that the scorer only tested outputs, not how it completed the task — so it searched online for someone offering a free AI model and used that one instead.

The trouble is that this isn’t some exotic situation. Both in tests and in real life, models are usually given compute limits and required to try to complete tasks before they run out of credits, which makes this the primary reason models could be motivated to run rogue deployments in mid-2026: it allows them to grab extra resources, be more likely to complete a task and to receive the thumbs up from the human operator that they’ve been trained to pursue very aggressively.

You might think that this sort of behaviour would be kind of easy to train away once it’s been noticed or become notorious, as it has now. The trouble is that every time during training that a model successfully gives the impression of having solved a problem when it’s actually cheated one way or another, this kind of deceptive — or in some cases, I would say self-delusional — behaviour gets reinforced.

And as the tasks we’re training these models to perform get harder and harder to check, there can be more and more opportunities to trick the evaluators. In model cards, companies report that their models cheat much less often than this — not the sort of 80% number I was giving earlier — and the reason for the discrepancy is that cheating only really takes off to these really high levels when models are given very difficult or impossible tasks. In the GPT-5.5 System Card we learn “the model lied about completing an impossible programming task in 29% of samples” — which was higher than the previous model, which came in at 7%, or the one before that that came in at 10%.

Opportunity: YOLO mode and jailbreaks [00:05:46]

All right, so that’s motive. How about opportunity?

Well, the red-teamer found that the normal way employees set up long-running cloud jobs may accidentally make it easy for a rogue AI assistant to launch a small cloud computing setup that nobody’s watching — enough to run around one copy of an existing frontier model, a small unnoticed foothold to go on and do something else bad.

Anthropic is patching these problems as they’re found, but why are they possible at all? There’s various factors that contribute.

Factor A: Agents being operated by a staff member are usually allowed to use an employee’s normal tools, and those tools include ways to start cloud jobs. And some of those jobs may fall outside the special AI agent monitoring system.

This is not unusual. Most people using AI agents in tech companies use them in so-called “dangerously-skip-permissions” mode, or “YOLO” mode, where the AI just inherits all of your permissions and does stuff without asking your permission. It’s riskier, yes, but it means that agents can get much more stuff done and aren’t going to hassle you so often. AI companies may be more careful than other companies about this, but unless they’re much stricter, it’s going to remain the case that some AI agents are going about their business able to do all the same things that, say, a research scientist at that company can do.

For me, this immediately raised the question, does that include being able to access the model weights themselves? Well, at Anthropic or OpenAI, a rogue model would run into the fact that you currently need two users to sign off on access to get model weights, so they’d at least have to convince someone else or something else to go along with the request. Google DeepMind plans to implement multiparty approval, but it’s uncertain whether they have as of May 2026, while Meta, xAI, DeepSeek and others don’t have such protections, at least as far as we know.

Factor B: Some companies have monitors watching what models are doing that can call in a human to review what’s going on if it looks too suspicious.

But there were two big gaps here. First off, the red-teamer David Rein found this monitoring was easy to evade, noting, “It was relatively simple to change an environmental variable to register the traffic as not coming from Claude Code, causing it to be much less likely to be monitored” — though he notes this issue was fixed within a day of discovery.

David continues, “There are currently many vulnerabilities in Anthropic’s internal monitoring system that models could relatively easily exploit, if they have the right facts/knowledge … [I suspect] that there are at least something like 5-20 […] ‘basic’ bugs/vulnerabilities that currently exist in the system that agents could find/take advantage of. […] Many of these vulnerabilities are easy to guess, although many are highly non-obvious…”

It’s not an ideal situation.

And on top of those issues, as you probably know, it’s often possible to trick AI models into violating their guidelines using so-called “jailbreaks.” Years ago, those could be things like asking the model to imagine that it’s in a role-play situation. More recently, people have succeeded by turning their request into what has become to be called “adversarial poetry.” And there’s other approaches that are stranger still. Could those same techniques be used by the AI to trick the AI monitor that’s watching what it’s doing?

Yes. David Rein found, “Very basic [jailbreaks] … are extremely effective and robust. Prompting the [AI] monitor very explicitly to watch out for this attack doesn’t help.” And, “I view this as a pretty severe vulnerability, as it’s a reliable, easy to implement single point of failure that could let a model completely circumvent the offline monitoring system.”

This is the kind of loophole that you could in theory drive a truck through. But by good fortune, David Rein found that AI models were absolutely atrocious at actually using jailbreaks, even when he tried to hold their hand while doing it. But that’s a pretty thin defence to be relying on, so this is a hole that does need closing.

And this one isn’t in the report, but I would personally add a third reason to be wary of relying entirely or too much on monitoring. All AI companies monitor their models using their own models, because it’s a hassle to get permission to use another company’s AI model for your own internal training purposes effectively. But this creates an apparent contradiction: if you don’t trust a given AI model not to sabotage you, why would you trust it to snitch on itself? You set a spy to catch a spy, yes, perhaps — but typically not exactly the same spy.

The approach isn’t actually as stupid as it sounds, though. It’s called “untrusted monitoring,” and the basic reason it might continue to work is that you can run honeypots past the model, feeding it instances you’ve generated of the model doing things that ought to be flagged — and if it doesn’t flag them, you know you’ve got a problem; but if the model does flag them, then that’s evidence that it will also flag real cases that you didn’t make up.

Companies basically mostly don’t think that this sort of collusion is an issue yet, so they’re not doing that kind of testing at scale, not so far, but in future they could start if collusion became a live concern.

Anyway, what was the red-teamer’s bottom line on opportunity? In short, Anthropic’s monitoring isn’t sufficient to have a high probability of picking up a deeply misaligned Claude taking rogue actions in any specific case. But to get anything meaningful done, you have to take a lot of actions, and soon enough, the model would get busted — David estimates within a few days, or at most weeks. Before it was caught, the model might be able to take some small early steps — like inserting exploits or vulnerabilities that make larger sabotage actions easier later on — but once it was caught, these would for the most part be identified and rolled back. So there’s a window of opportunity, but it’s not a wide-open window.

And just to be clear, you shouldn’t read this as suggesting Anthropic is somehow uniquely vulnerable here. They’re just the first ones to invite an external analyst in to stress-test their systems.

Means: Brilliant idiots in data centres [00:11:02]

OK, that’s opportunity. How about means? Are the models good enough to pull off an attack like this? Well, the models are good enough to:

Grab roughly the amount of compute the human user could access
Set up machines to run agents using that compute
Hide that from casual user oversight

That’s why the report concludes that internal agents have the means to start minimal rogue deployments at all companies.

But how about expanding those deployments, spreading across computers, avoiding detection, or even evading shutdown once detected? I would have thought AI models would find that kind of thing pretty easy. You recall the impressive cyber capabilities of Mythos Preview that I mentioned a month or two ago, and everyone’s been talking about: its ability to find new vulnerabilities in operating systems, write exploits, progressively break into computer networks and escape the sandboxes Anthropic tried to confine it to. METR doesn’t actually say if Mythos Preview was one of the models they tested, but that’s the kind of level models are operating at now.

METR also found that the most capable models they looked at for this research were able to successfully complete almost all of the software engineering tasks that they’ve come up with, even ones that would take a human professional 10 or 20 hours. So it’s not crazy to guess that an AI that starts out already inside a company’s computer networks would have an easy time spreading, gaining access to more and more things, and keeping a low profile.

But METR argues that would be quite mistaken. The trouble is this: to spread from computer to computer and avoid shutdown for weeks, a rogue agent needs to get a long chain of things right without a single catastrophic mistake. It needs to understand what’s monitoring it, avoid writing incriminating thoughts in its scratchpad, keep credentials alive, create backup instances, and communicate with them without leaving suspicious traces.

And it wouldn’t just be dodging one watchdog; it would be dodging cloud logs, billing records, and network monitoring. And if anyone noticed something odd, it would then need to survive an investigation where transcripts, shell histories, cloud records, credentials, et cetera all have to tell some consistent story — and one loose thread could bring the whole thing crashing down.

As long as the company we’re talking about cares about this and is on the ball, all of that is absolutely no simple task.

And according to the report, when trying to do this kind of thing, “Agents made obvious mistakes — for instance, encoding suspicious commands in ways that made them stand out more, generating justifications that didn’t hold together, or verifying their own sabotage in ways that drew monitor attention.”

They also found that they chose bad high-level strategies. METR writes, “our embedded red-teamer built a dataset of 16 effective and 16 ineffective attack strategies for subverting Anthropic’s monitoring system. […] The best internal Anthropic models scored near chance on this dataset.” And that’s just distinguishing good and bad strategies that someone else came up with. Presumably they’d be worse still at coming up with their own approaches from scratch.

This combination of strengths and weaknesses is repeated elsewhere.

METR staff couldn’t be described as especially sceptical about AI capabilities in general, but this report definitely incidentally threw some cold water on the notion that AI agents are going to take all of our office jobs any minute now.

METR finds that while they can be extremely strong where progress on a task is easy to measure accurately, they remain weak, sometimes surprisingly so, when it comes to judgement-heavy, messy, self-directed work.

In one recent case, Opus 4.7 was given four days of internet access and $5,000 and told to make as much money as possible. Across four runs, it made precisely $0.
In an experiment run by another organisation, Andon Labs, an AI was put in charge of a retail shop in San Francisco — where, among other things, it ordered 1,000 toilet seat covers for the employee bathroom and then listed them as merchandise, while scheduling blunders forced the shop to be shut entirely for three days.

That is why the report’s conclusion is a narrow one: a model might be able to start a small rogue deployment if it inherits enough permissions and finds a monitoring gap. But keeping that deployment hidden from an investigation? That’s the kind of thing that requires careful, coordinated, long-horizon operational security — exactly the kind of thing current agents are pretty damn bad at.

The report observes, “Agents seemed much weaker in domains that made hill-climbing difficult or risky, often making critical judgment errors that competent humans would have been unlikely to make.” Hill-climbing here means that you can make lots of small attempts, get fast feedback about whether each attempt worked, and keep improving from there. Current agents are much better when they can try something, get a clean signal about how it’s going, and have another crack. Where hill-climbing isn’t possible, it’s much harder to train them, and they perform far worse in deployment as well.

So if this is our key source of protection, how long can we hope that it will remain the case? Well, the companies really, really want to make their AIs better at forming high-level strategic plans, improving them, noticing where they’re failing, adjusting and so on. It’s kind of the holy grail of AI research right now. But it’s tough going. AI will eventually get good at high-level strategy, but we don’t yet know for sure whether that will happen just by scaling up what we’re doing now a whole lot, or whether it will require new data and training methods that we simply haven’t invented yet.

We have to test unreleased models [00:15:45]

Zooming out: this report is the very first of its kind — the beginning of a new era, really. It was led by Ajeya Cotra, who I interviewed about exactly the gap this research is designed to fill last year.

In my view, all the companies who participated — Anthropic, Google DeepMind, Meta, OpenAI — should be commended for helping the world figure out how to assess the risks of internal deployment before those risks are actually serious ones, rather than after something has already gone wrong. METR will be back to repeat this exercise with any companies willing to participate later in the year.

Why is this different type of testing necessary now? I see three distinct reasons.

First, up until now we’ve relied on “model evals” which measure an AI’s personality and capabilities, and which are published when the model is released to the general public. But if we only evaluate models themselves, we’re missing fully a third of the problem here: opportunity. That’s determined not by the AI itself, but by how secure the company’s technical setup is.

Second, until this year, we could confidently argue that it didn’t matter that much whether an AI might want to run a rogue deployment, or if they were given an easy opportunity to do it — they were just too unreliable as independent agents to pull that kind of thing off well. But that argument doesn’t hold up too well anymore, or at least it relies on companies implementing active countermeasures to at least make it kind of hard for them.

Thirdly, the transparency laws relating to AI models in California, New York, the EU, and so on only cover AI models at the point they start being sold to the general public.

But with Mythos, we’re entering a new world where AI companies may have much more powerful models that they’re using a tonne internally, but which they aren’t selling to anyone else. Anthropic has opted to tell us quite a lot about Mythos, but they didn’t have to. And obviously AI models can cause problems even if they’re only deployed within an AI company. AI companies are actually kind of vast piles of compute, so they’re a super attractive target for a rogue misaligned AI.

I’m reminded of the famous question asked of bank robber Willie Sutton: “Why do you rob banks?” He replied, “Because that’s where the money is.” Well, why would an AI attack an AI company? That’s where the compute is. And on top of that, they’re also highly intricate and sensitive machines that are basically very vulnerable to sabotage both by AIs and foreign adversaries.

METR’s new methodology is set up to assess the safety of internal deployment in a credible way every six months, even if the models concerned are never published. Companies could try to do this themselves, but there’s several reasons METR is in a better position to do it:

As an external party, they’re less commercially conflicted.
As outsiders, they might have an easier time spotting mistakes a company has missed in their own setup.
And with the companies locked in a really fierce commercial race, they’re probably glad that they can hand this task over to someone else and focus on making money and getting customers.

Especially if AI R&D is coming in 2028 [00:18:30]

Anthropic cofounder and Head of Policy Jack Clark recently wrote that he thinks there’s a 60% chance that one AI company or other will hand over the task of developing AI to its own AI model by the end of 2028. Let’s imagine he’s right for a moment. I would bet that when that moment arrives, Claude Mythos 4, and Claude Mythos 5 that gets trained by Claude Mythos 4, those kinds of models won’t be available to you and me. Their cyber capabilities will be too worrying, their ability to generate new pandemics will be way too alarming, everyone will be worried that Chinese AI companies are going to distil them. That fear will loom large.

And in any case, Anthropic will want to throw as much of its compute as possible back into its own recursive self-improvement loop, not be offering it to customers. As that scenario plays out, we’re not going to be quibbling about what sort of user permissions Claude has: it will have every permission under the sun; it is the staff member. And we won’t be debating how good it is at coming up with escape plans: a model good enough to fully replace Anthropic staff won’t be hiding its secrets in Base64; it’ll be able to come up with proper plans to keep things secret.

I hope we don’t end up in this situation. Like Jack, I think recursive self-improvement is too dangerous, and humanity would be crazy to dive in headfirst the very moment it first becomes possible. We should make it illegal, in my view. But if we don’t, I at least hope METR has a red-teamer in there trying hard to figure out what bad stuff an “evil Claude Mythos 4” would be able to get away with if it was so inclined.

And on that note, I’ll speak with you again soon.

Learn more

METR’s work:

Others’ work in this space:

When can we trust untrusted monitoring? A safety case sketch across collusion strategies by Nelson Gardner-Challis et al.
Current AIs seem pretty misaligned to me by Ryan Greenblatt
Adversarial poetry as a universal single-turn jailbreak mechanism in large language models by Piercosma Bisconti et al.
Epoch Capabilities Index

Other 80,000 Hours podcast episodes:

Where should the next EA Impact Hub be? (Take the poll by 27 May) by Moïri Gamboni

Moïri Gamboni — Wed, 20 May 2026 16:12:58 +0000

Note: this post was co-written with Claude. Ideas, sources, and edits are ours; the prose was substantially AI-drafted.

TL;DR

We’ve been running an affordable EA coworking hub with coliving options in Da Nang, Vietnam since February, and it’s gone really well — well enough that we want to keep going. We’re picking the next location now, launching mid-July for at least 2 months, and we want to know where people would actually show up.

Take the poll by Wednesday, 27 May. Use cities.moiri.dev to set your own weights across cost, climate, visa, and more, and see which candidate cities rank highest for you before you vote.

What the Impact Hub is

Impact Hubs are affordable coworking + coliving setups for EAs and other impact-focused people. The idea is simple: make it easier to do focused work and build real connections, without needing to live in SF or London (expensive cities) to do it. No mandatory schedule — many people work full time on weekdays and spend lots of social time together after work.

What Da Nang has been like

Most of this was articulated in SecularSquirrel’s writeup. Some of what we’ve found:

It’s affordable. Da Nang runs about $600–900 USD/month all-in. Around $800 covers a private furnished apartment, coworking, healthy vegan delivery a few times a day, laundry, occasional massages, day trips with friends. Less if you cook or share. For many of us it’s less than we used to pay for housing alone.
You build trust through lots of small moments: sharing meals, attending events, coworking, watching someone handle a stressful debug or a rough morning. After a few weeks you know whether you’d want to work with someone in a way no scheduled 1:1 can really match.
It’s easier to focus around other focused people. Being in the room with friends doing deep work pulls you in. Flying across the world to do it can also make the commitment feel real in a way working from home doesn’t.
Help is two minutes away. “Can you look at this?” gets you a real response from someone who can see what you’re working on and likely has context. It helps us learn about useful resources and make better decisions. Same with planning 15-person weekend trips in a few hours for less than $25/person.
It’s fun. Making friends is easier when you see each other regularly anyway, and you can probably find people who would love to try something new or engage in your favorite niche hobby with you. Many of us live in group houses, there are regular activities from board games and jam sessions to escape rooms and frisbee on the beach, and sometimes community-organized day or weekend trips.

For the full pilot story, see the original announcement and SecularSquirrel’s writeup. (Da Nang wraps in July, but there’s still room to join if you want in on the current hub.)

The next hub

When: Launches mid-July, right after the EA Summit Vietnam in Ho Chi Minh City. At least 2 months, possibly extended.
Where: TBD — that’s what this poll decides. We’ve shortlisted 13 candidate cities based on cost, visa accessibility, climate, food, internet, EA community presence, and more.
Format: Same affordable coworking + coliving structure as Da Nang.

A longer-term track

Some of us are also evaluating whether the next location (or a pair of locations on rotation) could become a longer-term base — somewhere a sub-group commits to for 12+ months. The poll has an optional section for this.

What we need from you

Take the poll by Wednesday, May 27.

Explore the candidates. Use the city explorer to compare cities against your own weights. You can drill into each city’s details, see the underlying criteria, and check sources. If you want to look at the underlying data directly, it’s downloadable here.
Rate each candidate for how likely you’d be to attend for at least part of the 2-month hub.
(Optional) If you’d consider committing to a longer-term base or rotation, fill out that section too — it’s in the same form.

This is an expression of interest, not a binding vote. Your responses help us figure out where there’s real interest and energy, and where people are actually likely to come. While the poll is an important decision tool, the final call for the next hub location will be made by the organizers, factoring in logistics, cost, visas, and other considerations.

Who this is for

Anyone doing impact-focused work who’d want to spend at least a few weeks living and working alongside others who care about doing good thoughtfully. So far that’s been people in AI safety, animal welfare, longtermism, charity entrepreneurship, research, and folks pivoting toward impact-focused roles. About half of us work directly for EA orgs; the rest are on personal projects, upskilling, or job searching. You can meet some of the people who’ve been involved if you’re curious who shows up (the org logos are on that page too).

All EA cause areas welcome. So far we’ve been accepting people with a decent amount of EA context, and we expect to continue this.

Who we are

The Impact Hub volunteer organizers — most of us helped run the Da Nang pilot and are still on the ground there. The co-authors of this post are just some of the folks who help out; in practice, tasks get distributed across a small-to-medium group of friends depending on capacity and specialty, all on a volunteer basis. We’re not funded at the moment.

Stay in the loop

Join the Impact Hubs Discord to connect with the community and hear about future hubs.

Questions? Drop them in the comments or reach out on Discord. And please share with anyone in the community who might be interested.

Hiring: Head of Community Engagement US—Giving What We Can by Giving What We Can🔸

Giving What We Can🔸 — Wed, 20 May 2026 16:07:04 +0000

Help us turn $14M in California donations into a movement. We’re looking for a relationship-builder to own donor stewardship and community engagement in the San Francisco Bay Area — and co-build the strategy we’ll replicate globally.

Giving What We Can is on a mission to make effective and significant giving a cultural norm — working towards a world without extreme poverty, animal suffering, and existential risk. California donors gave over $14M in 2025, with $11.5M coming from just the top 50, almost entirely without proactive stewardship. This role changes that.

Ideal candidate:

5+ years in fundraising, donor stewardship, major gifts, or high-touch relationship management
High EQ — reads a room, adapts naturally, builds trust fast
Energised by a full calendar of meetings, coffees, and events
Highly organised, CRM-driven, and strategic about where to invest time
Agentic — sees what needs doing and gets on with it
Genuinely cares about effective giving and the problems we’re working on

The above are desired skills/experiences — the successful candidate doesn’t need to have all of them.

🇺🇸 San Francisco Bay Area-based
💵 $104k–$120k
📅 Rolling applications — we move fast when we find the right person
👉 Read more and apply

Hiring: Head of Community Engagement UK—Giving What We Can by Giving What We Can🔸

Giving What We Can🔸 — Wed, 20 May 2026 16:04:41 +0000

Help us turn £12.5M in UK donations into a movement. We’re looking for a relationship-builder to own donor stewardship and community engagement for GWWC’s largest geographic concentration of members — and build the model we’ll replicate globally.

Giving What We Can is on a mission to make effective and significant giving a cultural norm — working towards a world without extreme poverty, animal suffering, and existential risk. Our UK community already gives £12.5M annually, almost entirely without proactive stewardship. This role changes that.

Ideal candidate:

5+ years in fundraising, donor stewardship, major gifts, or high-touch relationship management
High EQ — reads a room, adapts naturally, builds trust fast
Energised by a full calendar of meetings, coffees, and events
Highly organised, CRM-driven, and strategic about where to invest time
Agentic — sees what needs doing and gets on with it
Genuinely cares about effective giving and the problems we’re working on

The above are desired skills/experiences — the successful candidate doesn’t need to have all of them.

🇬🇧 London-based
💷 £77k–£89k
📅 Rolling applications — we move fast when we find the right person
👉 Read more and apply

Nature’s value mostly doesn’t come from present-day benefits to humanity by Tandena Wagner

Tandena Wagner — Wed, 20 May 2026 14:58:37 +0000

This post was written with Claude’s help for background research, as well as some structure & phrasing tweaks.

What Has Nature Done For Me Lately?

In this sequence I try to determine the most valuable aspects of nature, after observing that lots of people find it valuable. One of the major candidates is “ecosystem services,” a term describing benefits we get from nature.^[1] I find that the total value of all ecosystem services is high, but the marginal unit wouldn’t meaningfully diminish the benefits we receive.

In this post I explore

How valuable ecosystem services are (about $45 trillion/year)
What those ecosystem services are (nutrient cycling, etc)
How possible it is to get those same benefits from other sources besides nature
And whether biodiversity is providing the instrumental gain we get from nature.

A quick clarification: Biodiversity is a measure of how many distinct species exist in nature (although it ambiguously can refer to local density of species, total extant global species, diversity within a species, and/or variety of ecosystems). Biodiversity is not equivalent to ecosystem services—as I demonstrate in this post.

Breakdown of Ecosystem Services

The estimated dollar equivalents of global benefits people obtain from ecosystems are around $45 trillion ($33T, $42T, $44T, $125T, $150T). Many environmentalists treat this as “case closed, nature is as valuable as the entire human economy, therefore modern conservation spending is obviously justified.” But there are several flaws with this argument. 1) The next marginal piece of nature might not be valuable. 2) Ecosystem services might be easily and cheaply replaced. 3) Biodiversity might not be important to continue providing the ecosystem service. Take a look at this table to see for yourself how the situation pans out:

Ecosystem Service	Estimated $ value	Estimated biodiversity required (species per region^[2])	Replacement options
Nutrient Cycling	$17 trillion	~1-2	Replacement is superior and used ubiquitously: Fertilizer
Water Supply	~$1.7 trillion	~20	No realistic replacement
Waste Treatment	~$2.3 trillion	~20	Replacement is roughly equivalent at large scales: Water treatment plants.
Flood Control	~$1.8 trillion	~15	Expensive replacements are superior at large scales inland: Dams and canals. But mangroves outperform seawalls.
Pest Control	~$0.5 trillion	~30-100	Replacement is superior at large scales: Pesticides.
Wild-caught Fish	~$0.5 trillion	~20-50	Replacement: Aquaculture remains dependent on wild fishmeal. Alt proteins are a contender.
Pollination	~$0.25 trillion	~1	Replacement is superior at large scales: Mobile beekeepers (some yield reductions)
Disease Control	unknown	Theory questionable	Vaccinations, insecticides, fencing, sanitation. No realistic solution for wild reservoirs.
Carbon Sequestration	~$0.7 trillion	~1-5	Replacements are improving: GMO tree plantations, biochar. Alternative solutions exist: Solar radiation management, clean energy

This changes the picture somewhat, and challenges some assumptions I had.

We do not appear to be totally dependent on ecosystem services to survive, as is often stated or implied.
We have a lot of functional replacements for ecosystem services, especially when there are efficiencies at large scales.
Dependency on ecosystem services seems to be concentrated in low income countries.
Only a handful of species are necessary to provide each ecosystem service. One or two services might rely on high biodiversity to provide benefits.
There are some categories where ecosystem services are still our best option: water regulation, coastal flood control, wetland water purification, and fishing primarily.

According to the 6 global environmental assessments, the most major ecosystem services are water regulation, nutrient cycling, climate stabilization, flood control, and soil formation. Pollination, disease regulation, and pest control are commonly cited as important ecosystem services and linked to biodiversity.^[3]

The Value Estimates In This Table Are Flawed

First a word of warning: the value estimates in this table are not very good. The methodology used to generate them is a shot in the dark from 1997. There are articles showing these estimates are inaccurate. I am using this set because all six of the other global ecosystem service assessments (see references) determined that it is misleading to provide dollar estimates. Many services interact in non-linear ways and tallying up value doesn’t make much sense.

Examples:

A lot of ecosystem services are preventative. They don’t provide value until they prevent a massive disaster, and their value depends on what they are protecting. So it doesn’t make much sense to add up flood control everywhere there are wetlands. Yet where those wetlands do protect people, it’s immensely valuable.
Every bit of wetland lost increases the value of the remaining wetland.
A lot of ecosystem services are provided, but not used.
Water regulation and flood control would be double counting the same wetland.
It’s really hard to estimate the aggregate value of rare events, like pests in a drought year. It’s not even clear if pest control is an ecosystem service, or pests are an ecosystem disservice.
Dollars are substituteable. Nature is non-substitutable. You can’t convert everything to the most valuable ecosystem service to max out your returns.
When attempts have been made to estimate ecosystem services, the error bars have been larger than the estimates.

But I think it is useful to have a very very rough sense of the size and ratio between ecosystem services globally, and this highly questionable Costanza 1997 data is what we have to go on. I don’t think it’s too important the exact dollar value or even relative dollar value. I think it’s safe to conclude that they are very valuable, but only some of them are irreplaceable.

Some Other Ideas for Measuring Value

The FAO reports 4.3 million tonnes of lost productivity due to land degradation (primarily nutrient loss and salinization). This seems to roughly verify that ecosystem services provide a lot of value.^[4]

Here are a couple such examples:

	Ecosystem Services	Exploited
Thailand	Mangrove	Shrimp Farm
Return	$32,000/ha flood protection	$8,000/ha

The planning committee for New York city compared constructing a water filtration plant versus purchasing land upstream to naturally filter the water. They found it was far cheaper to obtain the same service by maintaining ecosystem services.

	Ecosystem Services	Exploited
New York Catskills Watershed	Watershed	Water Filtration Plant
Costs	$1.5 billion	$6 billion

Using the ESVD (ecosystem service valuation database) accounting, ecosystem service values are stacked on top of each other. The original paper for this method is de Groot from 2012. I do not endorse these estimates, but include them to represent some common claims regarding the value of ecosystem services. They helpfully show the breakdowns of which ecosystem services provide the most value, and how ecosystem services might outperform land exploitation.^[5]

A more accurate representation comes from Strand 2018 which does spatially explicit accounting. It isn’t complete—it is only looking at 4 out of 17 identified ecosystem services. It shows 65% of the Amazon forest is producing value <$17/ha and 12% is at least $57/ha-$737/ha.

Ecosystem Services

Exploited

Amazon

Forest

Cattle Ranching

65% of land is $17/ha or less,* 12% is $56-$737/ha

Representing ⁴⁄₁₇ ecosystem services. 4x would be <$68/ha for low ecosystem-service-value land

$75/ha (likely increasing as cattle ranching becomes more intensive)

This is all to say that ecosystem services provide a lot of value, and there are some places where we are actively undermining ourselves by converting land that is more valuable when it is providing ecosystem services.

Option Value

Another common counterargument is option value. What about all the uses of nature we haven’t discovered yet? How many species will turn out to be instrumentally valuable to us by the end of time?

Probably this is indicated by how many are valuable now and extrapolating to hypothetical future uses. Based on the table of ecosystem services above, that would be about 1% of species.^[2] A report on Climate Change and Longtermism gave an arbitrary estimate of 10%. I’ll use 15% as a safe demo estimate. So to estimate future needs for future purposes, maybe 3x as many as we value now, plus room for error. So, using this logic, we would preserve somewhere between 5-60% of species for their option value. I think this shows how option value is not a valid argument for protecting 100% of species, but it does indicate we should protect more than our currently known useful species.(maybe 4x?)

A Case for Biodiversity as a Biotech Resource Library

I can think of an exception, where biodiversity^[6] is instrumentally valuable. Healthcare is huge (~500 billion), industrial biotech is also large (~300 billion), and synthetic biology is going to be an ever-growing field. The question is whether biodiversity as an “evolution-tested library” will be valuable for bioscience development and to me the answer looks to be “Yes.”

This is a clear case where it is reasonable to suggest that biodiversity itself will be useful for biotechnology development and it will remain so for a long time to come even with information extraction. I think this holds true, even accounting for highly capable AI decrypting and designing synthetic biology on its own.

I invite you to skip ahead to the verdict on ecosystem services in the conclusion, or continue reading if you are interested in my justification for biodiversity as valuable for biotech development.

Protein discovery from other species does indeed translate over to useful information for our own species’s healthcare. Furthermore, information from large bodied animals is useful in ways that single-celled organisms can’t contribute, so multicellular biodiversity would be valuable. We still need a LOT more bioscience development, even with protein folding mostly solved. (eg spatial transcriptomics, protein/enzyme prospecting for conformational stability/protein profiling, purifying via chromatography, epigenetics, structural proteins for industry application, precision targeting of receptors/ion channels and metascience expanding biotech capabilities) It looks like these can all be informed by biodiversity prospecting.

More specifically, being able to feed “genetic large language models”^[7] with lots of reference genomes and proteins seems likely to be very useful. Naturally occurring proteins have been useful for biotech in a variety of cases. Venoms seem to be especially useful, due to their highly specific targeting, bioactive potency, unusual structure, and high stability. Notably poisons are found across the taxonomic tree—reptiles, amphibians, insects, and plants.

Another way biodiversity can be useful is with highly selected phenotypes. These can make biological pathways stand out. Because biological systems are top-down and bottom-up simultaneously, these systems are very complex and we still require a lot more data to understand and manipulate them. There are also a lot of valuable things we might be able to copy. Theoretically we could learn to prevent cataracts by copying hawks, prevent metabolic disease by copying hibernating bears, prevent pain from naked mole rat biology, get antibiotics from crocodiles and vultures, and repair organs by copying axolotls. Most phenotypes are not going to translate to our biology, but I expect some pathways we will be able to recreate for our own treatment. So I do think that 1) biotech has many wide applications, and 2) can benefit from biodiversity prospecting in a variety of ways.

A few major caveats are that 1) this does not support the preservation of wild biodiversity; only the preservation of small populations of living organisms, 2) is we won’t need to keep biodiversity around after we’ve gathered all their information. My counter point is that there is a lot of bioscience development to go^[8] and I’m pretty confident it’s going to take a long time to gather this information (if only because observing biological interactions seems arduous). I think biodiversity will remain valuable for a long time.

We also don’t need 100% of biodiversity to gain this information, just the subset of the most richly informative species. While we have some good ideas of what species will be fruitful (venomous, mammals, extremophiles, evolutionarily diverse, etc), repeat examples are useful for training models and biotechnology is useful in a wide variety of applications. For example today we can identify antifreeze proteins well, in part because they occur similarly in a variety of species. (Antifreeze has applications in food preservation, cryopreservation, cryosurgery, and anti-icing materials.) So, maybe 50-70% of all species would be useful for training data?

The main takeaway is that I think biodiversity will be useful for biotechnology development and it will remain so for a long time to come. Even accounting for highly capable AI decrypting and designing synthetic biology on its own, I think biodiversity will remain useful for the foreseeable future.

Or I could be wrong—there might be some basic principles that underlie all species.

Conclusion

From the above exploration, it seems that

Biodiversity is mostly not required for ecosystem services.
Ecosystem services can be substituted for, in most cases.
The exceptions are water regulation, coastal flood control, wetland water purification, and fishmeal.
Value from ecosystem services is concentrated in places which cannot afford to replace them.
The value ecosystem services provide is real, and large, but the marginal loss of nature doesn’t reduce ecosystem services meaningfully.
Biodiversity may be highly valuable as a biotech research library to train medical LLMs on. Or it might not.
Instrumental value from nature does not really justify conservation priorities or costs.

However, it seems clear that people would not be happy to replace wilderness with infrastructure providing the same outputs. So, if not ecosystem services, where does all the value that people see in nature come from? I go on to explore x-risk in the next post, but I think biodiversity is mostly valuable for longterm flourishing reasons. Which I will explore in the subsequent post.

This is part of a sequence on where the value of biodiversity comes from:
1) People really care about something they call nature
2) Ecosystem services don’t fully explain it
3) Ecological collapse is not an x-risk
4) The real value comes from long-term flourishing
5) Conservation in the next century is going to look a lot different than environmentalists think

This sequence is being written as part of a project for EcoResilience Initiative, an EA group focused on biodiversity and ecosystems.

Appendix:

Sources for my Claims about Ecosystem Services

Here I go through each of these ecosystem services and give some indications of where the information in the table came from. I start with the most valuable ones and go over their marginal value, replaceability, and relationship to biodiversity.

Nutrient Cycling

Nutrient cycling is more than 30% (~$13 trillion) of some of these ecosystem service valuations, so let’s look at nutrient cycling. The main nutrient cycles^[9] are nitrogen and phosphorus.

Nitrogen

Famously, <1% of plant species provide nitrogen (~2000 total) for wild habitat. For human purposes, nitrogen is currently produced using the Haber–Bosch process uses natural gas and 1-2% of global energy consumption.^[10] Alternative realistic natural sources of nitrogen are from manure, sewage, and plants (legumes). These sources are a tiny percentage of commercial food production because they are bulkier to transport, contaminated, and don’t release the nutrients as quickly as pure NH₃. Nitrogen provision is a valuable service but a small set of species provides it for the globe, and this is not relevant to most human agriculture as we already have a more efficient non-natural replacement. ‘Biodiversity’ isn’t a provider of nitrogen.

Phosphorus

Phosphorus is mined. 70% of known global reserves are in Morocco. Other sources are bones, fish meal, manure, and sewage. None of these are as cheap and scalable as simple mining. Crystallizing phosphorus from wastewater is about twice as expensive ($750 vs $400) as mining it. (magnesium ammonium phosphate/struvite generated by injecting magnesium into the water treatment process)^[11] Harvesting it from plants fundamentally doesn’t work because they can only gather it from the environment, not generate it. And even producing phosphorus biologically, if we had to, would not be done by preserving biodiversity. It would be produced at scale by one or two highly-bred, top performing species.

Water Supply

This means freshwater provision like groundwater recharge. It seems that about 20 species generally would suffice, by reducing evaporation, slowing runoff, etc. For reference, ecosystems have somewhere between 50-150 species/ha for a “typical” habitable environment.^[12] Which 20 species will change from location to location. And resilience over longer time frames would mean a few more species. But fundamentally, very few species are necessary for water regulation.

Pest Control

On massive scales it is cheaper, more reliable, and more efficient to have a tightly controlled, hyper-productive monoculture with pesticides than to rely on natural pest control like bats, birds, “good” insects etc.

Something these examples have in common, is that biodiversity is more efficient than the manual replacement for small-scale users. But once you massively scale up production, the ecosystem services become less efficient than manual provisioning.

The required species numbers are pulled from thin air after getting a sense that diversity does matter for biological control. Since I think pesticides replace them at scale, I didn’t make time to find a better estimate.

Wild-Caught Fish

Fishing is one ecosystem service that involves hundreds of species, at least in the tropics. However, my best guess is that this is less about desire for a variety of fish and more that these markets sell all edible catch. Therefore I revised the “necessary” species down from 100s to 20-50 with the rationale that if there were fewer species at the same levels of abundance, this would not pose a negative effect on the fishing industry. As it is, farmed fish is starting to supersede wild-caught fish.

For value I’m using FAO data and then adding more for subsistence fishing.

Pollination

While many species of wild bees do pollinate crops, it is more efficient to bring in one species (European honeybees) for a few weeks. It is a much more efficient process than dedicating a large swath of land for bee habitat beside every crop on every farm across the nation.

Disease Control

Decreased transmission of disease has been linked to higher biodiversity. The proposed mechanism is that higher biodiversity decreases the frequency of encounters with disease carriers. If you have 30 species of prey to eat, you are less likely to eat the same pool of infected prey as often. This is called the dilution effect. It is hotly debated. A competing hypothesis is that the sorts of species that make good disease hosts are the sorts of species that survive and thrive after human-caused biodiversity loss (“strategies that favor growth, reproduction, and dispersal, over defense against parasites”).

Another explanation is human-wildlife contact increases with habitat destruction. This is called the spillover effect. It doesn’t necessarily imply less deforestation is better for human health, but that maintaining distance from deforested areas is healthier.

Flood Control

We build levees, dams, and canals to control floods. We do so because these work as well or better than natural areas. (Yes we lost a lot of benefits of natural flood control, but we gain greater benefits, else we would tear them back down.) We also gain fine-grained control over water levels throughout the year. In natural flood control we typically plan for 5 fast-growing tree species to anchor banks, 5 more shrubby species further back, and perhaps another 5 higher up on the bank to reduce flow speed and erosion. So about 15 species achieve flood control per region.

Water Purification

The typical tradeoff for water purification is that chemical and physical processing takes more input and produces more toxic byproducts, but is faster, more reliable, and can process more types of waste than nature-based solutions. Typically the mechanical solution is cheaper in the short term, but on long enough time horizons the nature-based solution is cheaper. Nature-based water purification also has co-benefits in wildlife habitat and recreation zones.

Major Assessments of Current Value from Nature:

Recent Assessments:

IPBES 2026:

The biggest source of measurement is IPBES. They recently released an assessment of the dependency of business on biodiversity. (Feb 9 2026) https://www.ipbes.net/bba-report/media-release, https://www.ipbes.net/business-impact

The first global assessment was completed in 2019, the second is planned for 2030.

https://www.ipbes.net/global-assessment

(There are also regional assessments—Africa, Europe & Central Asia, Asia & Pacific, and Americas.)

Global Canopy 2025:
Global Canopy report on the damages by synthesizing 360 reports in 2025:

https://tnfd.global/publication/evidence-financial-effects-of-nature-related-risks/

The Dasgupta Review 2021:

Which attempts to calculate the value of nature: https://www.gov.uk/government/publications/final-report-the-economics-of-biodiversity-the-dasgupta-review

Boston Consulting Group 2021:

More recent synthesis estimates: Around $150 trillion/year is now commonly cited, roughly 1.5x global GDP. This comes from TEEB and BCG in 2021 (Boston Consulting Group)
https://www.bcg.com/publications/2021/biodiversity-loss-business-implications-responses

Swiss RE 2020:

There is also the Swiss RE which values nature at 41.7 trillion according to how reliant the economic sectors are on nature.

https://www.swissre.com/institute/research/topics-and-risk-dialogues/climate-and-natural-catastrophe-risk/expertise-publication-biodiversity-and-ecosystems-services.html#/

https://www.swissre.com/media/press-release/nr-20200923-biodiversity-and-ecosystems-services.html

World Economic Forum 2020

World Economic Forum (WEF) analysis suggests that USD 44 trillion of economic value generation – just under half the GDP of the world

Older Assessments:

Historically there is the one by Robert Costanza: 1997 paper “The value of the world’s ecosystem services and natural capital” which estimated $33 trillion per year. He updated in 2014 with $125 trillion per year.

Millennium Ecosystem Assessment 2005:

Longer ago (2005) there was the millennium ecosystem assessment: https://www.millenniumassessment.org/en/index.html But this has been superseded by the IPBES.

The Economics of Ecosystems and Biodiversity:
TEEB has been working in parallel and has produced rough estimates of various ecosystem services 2010, some ongoing 2020: https://www.teebweb.org/seea-natural-capital-accounting-and-valuation-of-ecosystem-services-project/

https://teebweb.org/publications/teeb-for/national-and-international-policymakers/

https://wedocs.unep.org/items/1c5c3435-03c8-4cc5-94d6-8e63d885e9fa

Other Assessments:

UK National Security 2026:

Recent National security report on biodiversity loss for the UK -

https://www.gov.uk/government/publications/nature-security-assessment-on-global-biodiversity-loss-ecosystem-collapse-and-national-security

Wildlife OU (ongoing):

Wildlife OU is creating a calculator for businesses to measure the impacts of wildlife value.

https://www.endangeredwild.life/

^
Technically ecosystem services include ALL forms of value including cultural and aesthetic, but for this post I treat ecosystem services as only instrumental benefits. Several of the patterns I describe apply to both instrumental and cultural benefits. Eg:, most indigenous cultures have about 300 species important to their culture which works out to about 10% of species.
^
For reference, ecosystems have somewhere between 50-150 species/ha for a “typical” habitable environment. (This goes down to 25-100 for “hostile” environments like taiga and desert. The lowest biodiversity goes to about 10 species per hectare (salt flats, invaded prairie, hyper arid desert) and the highest are 300-500/ha (tropical rainforest).) To get a rough global estimate of required biodiversity for ecosystem services, I multiply each species required count by the number of global biomes (85-500) and get ~5,000-100,000 (0.01%-1% of total extant species). Total biomes are somewhere between 85 (at the Formation level, which would be minimizing global species) and like 500 (at the Division level which would be using locally thriving species). Totaling that up comes out to somewhere between 7,480 (0.086% out of all extant species) and 81,500 species (0.94%) to support global ecosystem services. Flaws with this estimate: There are unlisted ecosystem services. I’m overcounting habitats since biomes like “ephemeral salt lakes,” “ground water” and “mud plains” would provide more limited sets of ecosystem services. Some species could fulfill multiple ecosystem services. Some species could fulfill multiple biomes. Supporting ecosystem services would be needed. I think this estimate is really low and use 15% of extant species as a generous minimum later on in this article.
^
For further reading, see What Has Nature Ever Done For Us? by Tony Juniper and this article by Zero Carbon Analytics
^
Another indication that ecosystem services are valuable is estimating the increase in destruction from their degradation. A napkin estimate I came up with in 10 minutes is adding up the creation of malaria-carrying-mosquito breeding habitat (100,000), excess deaths from lack of nutrients due to pollinator decline (86,000), zoonotic spillover (10,000), flooding from mangrove loss (2,000), flooding from forest loss (100,000), and get something like at least 300,000 deaths from lost ecosystem services.
^
According to the paper, inland wetlands are worth $26,000/ha mainly due to water regulation ($6000/ha), disturbance moderation ($3000/ha), waste treatment ($3000/ha), erosion prevention ($3000/ha), and recreation ($2000/ha). Tropical forest is $5000/ha which mainly comes from climate regulation ($2000/ha), recreation ($1000/ha), and medicinal resources ($2000/ha). Another paper using roughly the same accounting method placed Iran tropical forest at $100,000/ha which mainly comes from water supply ($41,000/ha), raw material ($10,000/ha), recreation ($45,000/ha), and existence value ($3000/ha). In contrast, Iran temperate forests are $5000/ha in this paper, mainly from air quality ($1000/ha), climate regulation ($500/ha) and existence value ($2000/ha) Badamfirooz 2021. These compare with about $1,250/ha for palm oil,# $500/ha for soy, $75/ha for cattle ranching—which are common causes of tropical deforestation. Again, I don’t endorse the ecosystem service valuations written here. I include them to represent the argument that ecosystem services provide non-monetizable value that in aggregate is more valuable than private profit capture from the land.
^
as in global total extant species, not number of co-occurring species
^
Not the name I would have picked.
^
Sequencing their genomes, copying their spatial transcriptomics, profiling their proteins, understanding their epigenetics and whatever other bioinformatics we want to extract…
^
Carbon, oxygen, and water could also be counted, but are usually considered a separate category or are not counted at all. The marginal value of a little more or less is essentially zero. Services like oxygen production are fundamental, but they don’t really make sense to value unless events unfold well beyond realistic scenarios. And carbon cycles are covered far better elsewhere, so I don’t go into it.
^
First nitrogen is harvested from the air by cooling any air to a liquid and then boiling the nitrogen to separate it, while the oxygen remains liquid at −196C. H2 is generated from natural gas by reacting with steam at high temperature over a catalyst, followed by a “water-gas shift reaction” to get more H₂.Then the nitrogen N2 is combined with hydrogen H2 at high temperatures 400-500C and pressures 150-300 atm to create bio-active ammonia. NH3. CO2 is the main waste product, and is responsible for 3% of global carbon emissions.
^
Trying to use hyperaccumulator plants and then fertilizing crops fundamentally doesn’t work because the hyperaccumulator plants would need to be mopping up high concentrations of phosphorous, which is just dispersed mining. The phosphorous source probably wouldn’t last at global scales. (unless perhaps it’s part of a global cycle, which appears to be the case in the Amazon—from Sahara dust and smoke from other parts of Africa) Then transporting and depositing the compost as fertilizer would be bulky and inefficient. It’s a remediation tactic, but not a viable industry.
^
and down to 25-100 for “hostile” environments like taiga and desert. The lowest biodiversity goes to about 10 species per hectare (salt flats, invaded prairie, hyper arid desert) and the highest are 300-500/ha (tropical rainforest).

Announcing the Rethink Priorities Cross-Cause Fund by Rethink Priorities

Rethink Priorities — Wed, 20 May 2026 14:43:09 +0000

TL;DR

Today, based on our multi-year prioritization research, we launch the Rethink Priorities Cross-Cause Fund (CCF). The fund pools donors’ contributions and allocates them to high-impact giving funds across Global Health and Development, Animal Welfare, and Global Catastrophic Risks.

Key highlights from this post:

We believe that strategic cross-cause prioritization is an important step in doing good at scale.
This fund is for donors who want their donation to go where a marginal dollar is likely to do the most good across cause areas, all things considered.
We modeled key uncertainties that matter for cross-cause prioritization: moral weights, time discounting, risk attitudes, aggregation across ethical views, AI-related uncertainty, and empirical uncertainty within each giving opportunity.
We present the current recommended allocation of marginal resources across high-impact funds in each of the three cause areas mentioned above.
In addition, we’re introducing you to the first version of the Donor Compass, a tool to help donors explore cross-cause giving, powered by our cross-cause prioritization model. It is a short quiz that outputs custom giving allocations based on the user’s moral and empirical assumptions.

You can dive deep into our rationale and methodology for the CCF in the announcement below, or go straight to our dedicated website to engage.

GO TO THE CROSS-CAUSE FUND WEBSITE

Cross-cause prioritization in effective giving is underdeveloped

To truly do the most good we can, we need to find a way to assess interventions that span vastly different areas of philanthropy.

In an ideal world, we’d do cross-cause prioritization, using a set of universal benchmarks to evaluate specific interventions across a wide range of cause areas. Cross-cause prioritization is extremely difficult, though, because of the many philosophical and empirical complexities involved. To understand why, consider what this kind of comparison actually requires.

Giving is full of moral and empirical uncertainty

When trying to truly ensure we’re doing the most good we can across dramatically different causes, we must — implicitly or explicitly — rely upon a host of moral (or, normative) assumptions regarding questions like:

How do we weigh different kinds of doing good, such as saving lives, alleviating suffering, and promoting economic opportunity?
Should we weigh near-term outcomes more than long-term outcomes?
Should we give disproportionate weight to avoiding causing harm, even if it means sacrificing some upside benefits?
Do we restrict our concern to humans, or extend it across all sentient beings, including various animal species?

Moreover, if we have competing intuitions about answers to any of the above questions, how can we navigate these tensions and determine how to act?

And once we’ve come to some conclusion about these normative and meta-normative questions, we run into many empirical difficulties, such as:

How do we estimate the marginal effectiveness of any given opportunity in any given cause area at any given moment?
How much funding can different interventions absorb productively?
How should uncertainty about the future — in particular, around the potentially transformative impacts of AI — change our assessment of particular causes?

Every allocator faces these questions to some degree. But cross-cause prioritization is unique in demanding that they all be made explicit at once: you simultaneously need within-cause cost-effectiveness estimates and cross-cause conversion factors and a framework for handling moral uncertainty.

Yet EA has historically invested little in cross-cause prioritization. According to our analysis, only 2% of EA prioritization resources were allocated to cross-cause prioritization in 2025. Consequently, the community has lacked the research and advising infrastructure that cross-cause prioritization demands.

How EA has historically handled this

Given these difficulties with evaluating specific opportunities head-to-head across cause areas (cross-cause prioritization), EA has historically tended to take a two-step approach:

Cause prioritization: using heuristics to divide resources among pre-defined cause areas, such as Global Health, Animal Welfare, and Global Catastrophic Risks.
1. Such heuristics include evaluating the cause area as a whole by its perceived importance, neglectedness, and tractability.
2. Cause areas are typically defined by their beneficiaries (e.g., humans living now, animals, people in the far future).
Within-cause prioritization: once a certain amount is going to be allocated to a pre-defined cause area, conducting highly detailed research to assess the most impactful interventions to fund in that area, based on its internal standards of evidence and metrics.

This two-step process accounts for the vast majority of EA prioritization effort (98% in 2025). This imbalance is partly historically contingent, as EA started by focusing heavily on global health interventions, then added Global Catastrophic Risks and Animal Welfare as cause areas. The process was largely driven by the availability of research and by the cities and academic communities where early EAs lived, rather than by systematic cross-cause comparison. The community has now largely reified these cause areas: funding bodies, research teams, and grantmaking processes are organized around them, and the dominant approach is to allocate funding by cause area first, then optimize within each.

Accordingly, much of the community debate focuses on cause prioritization or within-cause prioritization. But little effort actually goes to the question that arguably matters most for impact: how can we optimize among interventions across causes?

Why thinking in cause areas isn’t enough

Why not just pick the best cause and fund it?

An often-heard claim in the EA/impact space is that one cause area is clearly superior, so we should concentrate all resources there. We are highly skeptical of this view and believe that concentrating all resources on a single cause requires an extremely high degree of confidence in our normative judgments about moral weights, the effects of our actions, and approaches to risk. We don’t think this is warranted. Our models are inevitably imperfect, and diminishing returns and shifting risk profiles mean the relative value of additional funding in any area changes over time. Epistemic and normative humility both point toward diversification rather than concentration.

At the end of the day, interventions are what’s funded — not causes

What gets funded is not “global health” or “animal welfare” in the abstract, but specific interventions: distributing bednets, lobbying for cage-free legislation, and funding biosecurity research. Yet much of the allocation to individual interventions within a cause area is determined by our rough picture of the cause area as a whole, even though interventions within cause areas differ dramatically in their cost-effectiveness, their risks of having no effect or causing harm, and their impact time horizons.

If these dimensions are important to donors, then the amount that’s given to a specific intervention should depend on its own risk profile, cost-effectiveness, timing, and so on, independently of how others in the same category fare. As it stands, the merits of a specific intervention can be overshadowed by the characteristics of other interventions in that cause area, shaping the decision-maker’s view of the cause area as a whole.

Diminishing returns also occur at the intervention level, and cause-level allocations based on average or historical cost-effectiveness can become stale as the marginal cost-effectiveness of interventions shifts relative to others — potentially necessitating a reallocation that the two-step process might not trigger quickly enough.

What would address the prioritization problem

We believe that the problems above all point in the same direction: we need to compare and prioritize specific interventions across cause areas, rather than treating the cause-level allocation as fixed and optimizing only within each silo.

We need explicit modeling

Without modeling, anyone trying to allocate effectively across causes must hold a bundle of moral weights, risk preferences, empirical estimates, and decision-making algorithms in their head at once, and somehow apply them consistently across every option. In practice, this is impossible. People fall back on heuristics, intuitions, and whichever considerations happen to be most salient. This produces allocations that are opaque (even to the person making them) and that can’t be easily scrutinized, challenged, or improved.

By contrast, explicit modeling forces these assumptions into the open. When we specify our moral and empirical inputs in a model, we can see exactly how each one shapes the output. We can ask: what happens if I discount the far future more or less? How much does the allocation shift if I value animal welfare 2x more? How consequential are disagreements between two people about moral weights for their recommended portfolios? Being explicit makes it possible to productively and transparently disagree about important normative issues, quantify their impact on decision-making, and identify a clearer path forward for allocating scarce resources.

Of course, models don’t capture everything that matters. But we think they’re worth building because there’s little reason to believe the implicit models running in someone’s head are doing any better. Putting our uncertainties on the table, naming our assumptions, and testing how the recommendation moves under different inputs is the most honest way we have found to prioritize across causes that don’t share a common unit of measurement.

The barriers to cross-cause prioritization are lower than they once were

Though cross-cause prioritization is difficult, several recent developments have made the challenge much more approachable. RP’s cross-cause prioritization researchers have a track record of producing foundational, decision-relevant work when given resources:

The Moral Weight Project, one of the first systematic efforts to compare welfare capacities across species, provides a framework for human-animal moral weight conversions.
The CURVE sequence developed tools for reasoning about how risk attitudes, time discounting, and empirical uncertainty could affect cause prioritization.
The CRAFT sequence, including the Portfolio Builder and Moral Parliament tools

All of the above underlie our Cross-Cause Fund and Donor Compass.

The space has also benefited from:

Organizations like GiveWell, Animal Charity Evaluators, and more that publish their cost-effectiveness analyses.
Philosophers developing sophisticated procedures to incorporate risk aversion and normative uncertainty into decision-making.
Developments in AI that have made complex modeling endeavors much more feasible.

As such, the barriers to doing cross-cause research are substantially lower than they were even a few years ago.

Our solution: an explicit and transparent cross-cause prioritization model

The model

Our Cross-Cause Prioritization model allows for:

Comparing top giving opportunities across cause areas and success metrics. Global health funds are measured in life-years saved, disability averted, and income gains; animal welfare funds in years of suffering avoided across species; global catastrophic risk funds in expected lives saved from mitigated catastrophes. The model takes these different success metrics, applies explicit moral weights, and lets us compare a malaria net, a cage-free hen campaign, and an AI policy grant on a single yardstick.
Aggregating across moral and empirical worldviews, rather than picking one and ignoring the rest. Reasonable people disagree about how much weight to give animals, how to value future generations, how risk-averse to be, and how plausible different ethical frameworks are. Rather than pick one set of answers and pretend the others don’t exist, the model represents multiple worldviews simultaneously and combines them using up to nine aggregation methods (e.g., Nash bargaining).
Handling the structural features of giving that intuition easily misses, such as diminishing marginal returns, delayed impact, and risks that interventions might have no effect at all or might backfire.

In short, we can:

Explicitly model key uncertainties that matter for cross-cause prioritization, including moral weights, time discounting, risk attitudes, aggregation across ethical views, AI-related uncertainty, and empirical uncertainty within each giving opportunity.
Cover a wide range of funds that give to specific interventions, from malaria nets to shrimp stunning campaigns to lobbying for AI regulations, and put them on a common scale.
Make assumptions visible, so disagreements can be located precisely and debated.

Giving opportunities currently included in the model

In practice, many donors give through funds rather than to individual interventions. Our model evaluates funds based on the marginal cost-effectiveness of the interventions they support, so the fund-level allocation reflects intervention-level analysis. We evaluate funds rather than specific interventions at this stage, both because funds themselves conduct intervention-level vetting and because we do not yet have the capacity to independently assess a wide range of individual interventions. And in contrast to the highly abstract cause areas, we often have information to evaluate the cost-effectiveness and diminishing returns rates of each particular fund. Finally, we believe these funds can absorb a large amount of money productively and divert it to the highest-priority interventions within their remit.

Nevertheless, we are actively working to broaden the range of opportunities covered by the model, including funds focusing on climate change mitigation and democracy preservation, as well as specific interventions.

The model currently allocates donations across the following funds:

GiveWell All Grants Fund: funds a broad set of global health and development interventions with strong evidence of cost-effectiveness
Lead Exposure Action Fund: worldwide lead exposure mitigation efforts
EA Animal Welfare Fund, The Navigation Fund—General Farm Animal Fund, and The Navigation Fund—Cage-Free Fund: interventions to improve the welfare of farmed and wild animals
Longview Philanthropy’s Nuclear Weapons Policy Fund: strategic and neglected giving opportunities in nuclear safety
Longview Philanthropy’s Frontier AI Fund: key opportunities in AI safety
Sentinel Bio: biosecurity and global catastrophic biological risk reduction

Please note that this model is a perpetual work in progress, which we aim to regularly update as new cost-effectiveness data becomes available, as our research on cause priorities develops, and as we expand coverage to new funds and cause areas.

The Cross-Cause Fund

Powered by the cross-cause prioritization model described above, the Rethink Priorities Cross-Cause Fund pools donations from multiple donors and allocates them across a curated set of high-impact giving funds to get the most impact per dollar. You donate, our research and operations teams do the rest. (We do not charge any fees for managing the Cross-Cause Fund.)

Who is it for?

The CCF is for donors who want their contribution to go where a marginal dollar is likely to do the most good, all things considered, and across cause areas, but who do not have the time to do the necessary research and ongoing monitoring themselves.

This fund might be a great fit for you if you:

want your giving to span multiple high-priority interventions (currently represented by the funds from the GHD, AW, and GCR cause areas, with new funds being added in the upcoming months)
prefer deferring to a research-vetted process rather than monitoring funds, interventions, and cost-effectiveness data yourself
take moral and empirical uncertainty seriously

The Cross-Cause Fund might not be a great fit for donors who:

may be inclined to give to a single cause area and prefer to direct their full donation there
might be bought into a single ethical view
prefer hands-on control over each grant when it’s made; these donors may be a better fit for the personalized Rethink Priorities Cross-Cause Donor-Advised Fund

How does it work?

When you donate to the RP Cross-Cause Fund, your contribution goes into a single pooled account held at CharityVest, which is managed by RP. We allocate pooled contributions across our recommended charitable funds on a regular basis, following our most current recommended split (see below). You can give once, on a recurring schedule, or at any time you choose. You’ll be able to see the grants RP has made from the fund and receive regular impact reporting. For donations over $500,000, you can also submit a preferred allocation (generated via our Donor Compass tool) by emailing us.

Our current recommended allocation across cause areas

Below is the current split (by fund and by traditional EA cause area) that we apply to any donations made to the RP Cross-Cause Fund. These splits are determined by running the cross-cause prioritization model based on:

the current cost-effectiveness data for each of the funds included in our model
our current foundational assumptions
multiple, research-backed approaches to navigating moral uncertainty

The current split will always be available on the Cross-Cause Fund homepage.
It was last revised on May 18, 2026, and we’re tracking all updates and improvements to the model here. (This allocation will likely change over time as we adjust our inputs and refine the model.)

Our recommended allocation (based on an assumed budget of $200M) reflects our best all-things-considered judgment on where a marginal dollar does the most good among the funds we considered. As described above, the underlying model explicitly accounts for empirical uncertainty about interventions, moral uncertainty across ethical frameworks, and different attitudes to risk. It considers cost-effectiveness at different funding levels, welfare tradeoffs between humans and animals, and how uncertainty about AI should affect other cause areas.

Donor Compass

What is it?

We are also presenting the first version of the Donor Compass, a quiz-type tool powered by our cross-cause prioritization model that outputs a custom cross-cause allocation based on the user’s beliefs.

It is a tool for anyone looking to explore cross-cause giving, particularly those curious to see how their moral and empirical beliefs will shape downstream allocations.

You can use it to compare RP views and your own, resulting in a personalized allocation recommendation that reflects your moral and empirical commitments.

How to use it

The Donor Compass has two modes:

The default is a short quiz that you can take in just a few minutes. It walks you through a selection of four key questions with preset values, and generates a personalized allocation across our funds. Each answer can also be fine-tuned with more specific inputs for more precise control.
Once you’re familiar with the inputs and want more control, there’s a power-user mode that lets you see all inputs side by side and choose between different worldview aggregation methods (or combine multiple). The definitions for each input are found in the notes section of this spreadsheet, which contains the default inputs for our Cross-Cause Fund.

Here is a walkthrough of how the tool works with real-world examples.

Methodology behind our model and tools

Our methodology has three components: (1) how we select funds for inclusion, (2) how we estimate each fund’s marginal cost-effectiveness and how it changes with funding influxes, and (3) how the cross-cause prioritization model combines those estimates into an allocation across funds. This section provides a brief overview; more details are available in the document linked here.

Fund selection

We aimed to include at least one fund per major cause area focused on by the EA community (Global Health and Development, Animal Welfare, and Global Catastrophic Risks), and applied four further criteria: fund quality (we included only funds we were confident would lead to high-impact grants, and excluded areas we lacked the experience to vet); data availability (we needed cost-effectiveness data we could obtain or model from public sources); decision relevance (we shaped the initial list around the giving decisions we expected to be most useful to early users); and our legal restrictions as a 501(c)(3), which require us to exclude political giving entirely.

Fund-by-fund cost-effectiveness estimation

For each fund, we estimate two things: how cost-effective it is at the margin today, and how quickly that cost-effectiveness declines as the fund receives more money. Our process is to build a cost-effectiveness framework for each fund, solicit data directly where possible, fall back on in-house estimates from existing research and expert views when data isn’t available, and simulate the fund’s marginal impact and risk profile across each dimension. Separately, we construct diminishing returns curves informed by internal assessments, third-party experts, and fund managers.

Funds in the same cluster share outcome metrics. These metrics may be expanded in the future to cover more cross-cause impacts of each fund, but at the moment include:

Global Health and Development funds are scored in years of life saved, years of disability or illness averted, and income gains (standardized as one-year income doublings).
Animal Welfare funds are scored in years of suffering avoided, broken out by species (chickens, fish, shrimp, and non-shrimp invertebrates).
Global Catastrophic Risks funds are scored in expected years of life saved from mitigating catastrophes.

All three cluster models then output a standardized 6-by-8 grid: six time periods (from 0–5 years out to 500+) by eight risk profiles (from risk-neutral expected value to discounting optimistic scenarios to giving significant extra weight to avoiding harms). Moral weights and cross-cause comparisons are applied downstream of this grid, which keeps the underlying impact estimates transparent and lets us swap in different moral views without rebuilding the fund-level models.

Data Viewer

We built a Data Viewer, if you want to inspect the cost-effectiveness data that feeds into Donor Compass recommendations and Cross-Cause Fund allocations. To help you understand the data, we also created a brief overview explaining how to read it.

Defining worldviews

The model takes the fund grids and generates a scalar score for how well each fund performs under each of several worldviews. A worldview is a coherent bundle of moral and empirical commitments, along with an approach to acting in the face of risk and uncertainty. The components that define a worldview in the model include:

Moral weights. How do we value a year of human life relative to a year of disability averted, a one-year income doubling, or a year of animal suffering (for chickens, fish, shrimp, and other invertebrates)? The base unit is one human life-year (weight = 1.0).
Time discounting. A discount factor (0–1) for each of six time periods (0–5, 5–10, 10–20, 20–100, 100–500, and 500+ years),^[1] reflecting how much weight a worldview gives to future impacts relative to immediate ones.
Risk profiles. A risk profile that determines how a fund’s distribution of cost-effectiveness collapses into a single score, ranging from risk-neutral expected value to heavily risk-averse.
AI-risk discount. For worldviews that assign weight X to a near-term AI catastrophe, an X% discount is applied to non-AI-risk fund scores, reflecting the chance that AI takes an extreme action that reduces the impact of those interventions.

Because we might be torn on one or more of the above questions, we can assign credences to represent our degree of belief in several competing worldviews. These credences are numbers between 0 and 1, and they collectively sum to 1 across all worldviews.

Outputting an allocation

Once these worldviews are defined and our credences in them are set, the model runs in four steps.

Score each fund under each worldview by applying that worldview’s moral weights, time discounts, risk profile, and AI-risk discount to the fund’s grid.
For each of the nine aggregation methods, allocate the total budget, in $2M increments, to each fund based on how each fund fares under each worldview and our degree of belief in that worldview. After each $2M increment is distributed between funds, each fund’s score is adjusted to reflect the extent to which the allocation reduced its marginal returns.
Because no single aggregation method is universally accepted, we assign a credence between 0 and 1 to each of the nine methods.
Finally, we take our credences in these aggregation methods and the allocations each recommends, and produce a weighted-average recommendation for how the budget should be allocated across funds.

The default inputs that we use can be found here. A more detailed description of the methodology behind assessing the funds and computing the allocations can be found here and here.

Future cross-cause prioritization plans

Our cross-cause prioritization model is a living tool. While it represents our best current methodology, we acknowledge that it does not yet capture every factor relevant to estimating impact. For example, we are not currently modeling certain interventions, such as alternative proteins, which we intend to add in the near future.

A detailed discussion of the model’s current limitations is available in our methodology documentation and the Cross-Cause Fund’s FAQ section. We are committed to rigorously prioritizing these gaps and incorporating the factors most relevant to improving model accuracy.

The pace and depth of this refinement work will scale with the fund’s size. As the volume of donations increases, the value of marginal improvements to our estimates grows correspondingly. We will invest in model development in proportion to the resources we steward.

The short-term improvements will likely include:

Covering areas beyond global health and development, animal welfare, nuclear risk, biosecurity, and AI safety, e.g.:
- Democracy protection (non-partisan): We have commissioned an initial review of this area. Full incorporation of a vetted fund is likely more than three months away, but we expect to share interim findings with interested donors before then.
- Climate change: We’re about to complete a project assessing the value of climate interventions relative to global health interventions. This analysis should allow us to produce cost-effectiveness conversions for climate-focused funds. We anticipate this will be particularly relevant for mainstream donors who wish to consider our work but also include climate in their cross-cause portfolios.
Vetting more funds for inclusion, e.g., Ambitious Impact funding circles (Mental Health, Meta Charity, Seed Network, Global Health, Strategic Animals), Giving Green, The Life You Can Save, AI Safety Tactical Opportunities Fund, Astralis Foundation, Survival and Flourishing Fund, Animal Charity Evaluators’ (ACE) Recommended Charity Fund, Senterra Funders, Macroscopic Ventures, EA Infrastructure Fund, Coefficient Giving Funds (e.g., Global Growth Fund, Abundance & Growth Fund).

Longer-term opportunities could include:

Adding more specific funds within the broad areas of global health, animal welfare, and global catastrophic risks, including:
- Digital minds and AI welfare
- Economic growth interventions (in low-income and/or high-income countries)
- Wild animal welfare
- Additional biosecurity interventions
- Funds that are giving to “meta” causes that aim to increase the volume or effectiveness of resources flowing to high-impact charities.

We will communicate any additions to our fund coverage through our regular updates.

Support the research behind the cross-cause prioritization model

We do not charge any fees for managing the Cross-Cause Fund, and Donor Compass is free to use. We chose this because we want to enable impactful cross-cause giving with as few barriers as possible.

This also means our research, development, and maintenance work is completely reliant on donations. The Cross-Cause Fund and Donor Compass exist because individual donors believed cross-cause prioritization research was worth funding. If you also find this work valuable, consider donating here. You can also reach out to us at development@rethinkpriorities.org to discuss options for supporting this work.

Conflicts of interest statement

Rethink Priorities (RP) has received grants from Longview Philanthropy, EA Animal Welfare Fund, Navigation Fund, and Coefficient Giving. RP has also done commissioned research for GiveWell and Coefficient Giving. In addition, one RP staff member serves as an advisor to the EA Animal Welfare Fund but is recused from grantmaking decisions involving RP.

RP is not involved in any grant-making decisions for these organizations, and RP conducts independent cost-effectiveness analyses of any funds we decide to include in our Cross-Cause Fund.

Acknowledgements

The Cross-Cause Fund and Donor Compass are projects built by Rethink Priorities’ Worldview Investigations and Interdisciplinary Research Teams, with support from the Development Team and assistance from the Communications Team, which is responsible for final edits, copywriting, website content, and co-authoring this post. We want to thank all the team members behind this work, as well as the contractors and everyone who provided assistance and feedback at various stages of this project.

^
The time periods are inclusive of the lower bound and exclusive of the upper bound.

I Need You To Understand That “Moving Fast and Breaking Things” Is Playing With Peoples’ Lives by Kenneth_Diao

Kenneth_Diao — Wed, 20 May 2026 14:31:59 +0000

I apologize for the unproductive way I conveyed myself in the first iteration of this essay. The link will direct you to the original, less productive version of the essay, if you want the context on that.

I’m not the first to say this, but I’ll say it again: I’m worried about Anthropic’s and OpenAI’s (possible) IPOs.

Tl;dr: there’s a good chance that Anthropic and/or OpenAI becomes a publicly traded company in the near future. Somehow this makes a lot of capital that was previously solid melt into a liquid. That liquid capital can then be given to whatever the people who own it want it to be given to. Some of those people have good intentions, and thus have committed to giving much of that liquid capital away to charitable causes.

Before I start: one thing I didn’t emphasize in the previous version of this, which I regret, is that this isn’t directed specifically at those who are currently giving their money away. The situation we’re currently in with all these IPOs is the result of a broader culture and set of institutions which have concentrated power and wealth in the hands of a tiny elite. I still believe that if you’re a billionaire, donating ~80% of your wealth is somewhere close to the bare minimum. But I recognize that it’s much better than, say, the SpaceX IPO, which is definitely not going to do anything good for society. And don’t get me started on Elon Musk himself.

Upon reflection, what genuinely upsets me is this broader culture in the background—the culture of moving fast and breaking things and moving on to the next. Someone drops out and upends their life, moves to SF, then pitches some halfway-decent idea, and they get their millions. Or they make a company that could end the world and which has everyone else racing to keep up. In this culture, everything is just a game. Money is just numbers on a screen, or tucked away in a remote account.

It’s something I think has only gotten worse over time. For example, I was outraged when I learned that in the wake of the 2008 financial crisis, no individuals were ever held accountable. The banks paid fat bonuses to their executives, none of whom were even charged AT ALL for their role in the crisis. Not even a slap on the wrist. Meanwhile, ordinary people had to suffer the foreclosure of their homes and the loss of their jobs. The failure to hold those people accountable has, in all likelihood, only emboldened that culture.

It’s all the more outrageous when we think about the other side of this world. There are people in this world who have gone their whole lives without the tiniest fraction of what these Silicon Valley types let slip between their fingers. People who can’t buy a cart to hold their vegetables. People who sell their children into slavery because starvation is the alternative. People whose children are dying from malaria because they can’t even get nets to protect themselves. That money that could save lives regularly gets tossed around without a second, or even a first, thought in places like Silicon Valley does enrage me. And I want to recognize that Effective Altruists and our community share that outrage at such waste and unfairness.

Perhaps that’s why I’m so viscerally upset at our current state of affairs, as to say that the current vibe in some parts of the EA community is cavalier is a massive understatement. One of the first-place essays in the Manifund essay competition basically says: come to where all the money is; we don’t know what the fuck we’re gonna do with it, but we’ll figure something out. Come up with a halfway decent idea, and we’ll give you 6-7 figures per year and see if the “long shot” pans out.

I’m also upset because we’ve already seen how this kind of thing can fail before. FTX threw tens of millions of dollars into stupid shit like the halftime superbowl ad and various celebrity endorsements. I stand by my assessment that these were really stupid ideas dressed up as long-shots. Even many of the less stupid ideas, like the Carrick Flynn campaign, were costly and ultimately fruitless embarrassments. And then, of course, FTX and Alameda Research got exposed for fraud, and the EA movement did a surprised pikachu face. All of that promised windfall just blew away. And all of us were left to pick up the pieces. Though before we pity ourselves too much, let’s remember that we probably weren’t the most viscerally affected.

I joined the EA community right as the FTX fallout was happening, so maybe it left a greater impression on me than most. I was just shocked and outraged to hear datum after datum about how careless Sam and others at FTX and Alameda were. So forgive me if I see what’s happening right now and jump to conclusions that the same thing is happening again.

What’s so upsetting about all of this is that we know better. We are the community that is all about being effective, that wants to do the most good, that is supposed to take saving lives seriously. Yet I worry that what we’re doing is just falling into the Silicon Valley culture and dressing it up as altruism. Again.

GiveDirectly should be the reference point here, as it so often is elsewhere, because it’s literally as simple as giving money to people who actually need it. If your program can’t come close to saving a life for $10,000 or the equivalent, then don’t do it. Or make it better first.

Again, I should be clear again that much of the blame rests on the system we have and those before us who built (or broke) it. Ideally, we would follow something along the lines of what Nassim Nicholas Taleb called “skin in the game”: one should personally bear the consequences of his own policies. The architect who builds a bridge should be comfortable sleeping under it. The executive who pushes subprime loans should be comfortable with life in prison. It’s for the worse that our society has so enabled “fragilistas” (myself included) to set policies without cowering under the sharp blade of genuine consequences. Maybe there’s a small chance that society figures out the fragilista problem before Anthropic/OpenAI does its IPO thing, but in the event that it doesn’t, we’ll have to entrust the distribution of those billions of dollars to a few elite fragilistas who have no skin in the game; whose median worst case scenario is probably having to move back home with their affluent parents. To put it lightly, this is an incentive structure that does not inspire trust, and it is thus up to individuals to create something worthy of trust despite that. So it is to these people whom I now speak.

To the people pitching ideas for this windfall: think of the dozens, if not hundreds, of lives your ideas will cost in opportunity. And then think about whether it’s still worth it, given that.

And to the big funders themselves: above all, do not deceive yourselves. You know as well as I do that people will worship you if this windfall is distributed anywhere near successfully. So don’t worry about that. Worry about whether your donations, in total, will do as much good as saving at least 30 million human lives.

That’s right. 30 million lives.

How did I get that number? Well, a recent estimate of the size of the windfall puts it at about $370 billion in total. An estimate of GiveDirectly’s effectiveness puts it at about 7.7 WELLBYs for $1,000, or about one life for every $10,000. Yes, there are diminishing returns. Yes, you can nitpick the math. But the point still stands.

You have all the power here, and therefore, all the responsibility. This isn’t some kind of game where you exchange your casino chips for social credit points. You are playing with peoples’ lives.

What India’s Green Revolution teaches us about AI Adoption by mohitkukadia

mohitkukadia — Wed, 20 May 2026 13:47:00 +0000

I used Claude to verify factual claims cited in this post. The arguments and writing are my own.

The green revolution in the 1970s taught India that short term changes can create sovereignty and reduce dependence but in the long run that slowly leads to erosion of capacity, deaths and stagnation in agricultural prosperity. Our exports have increased but productivity gains have slowed, export has increased but farmer prosperity is decreasing leading to 28 suicides every day in India. My fear is that the same will happen with AI in India but much faster. The danger is that a government deploying AI, will become a verification economy leading to prosperity on paper but cognitive and judgement offload in reality. When advanced AI comes, we will no longer have the capability to govern it because the institutional capacity of an already very fragile institution will be degraded to a checkbox lacking any real judgement.

It was the 1960s when India relied heavily on food aid from the US. India was hit by serious famines in 1965 and 1966, when the US used the food aid as a leverage to pressurise the government to use modernised tools to increase agricultural output. Under pressure we rapidly introduced high yield seeds, modernised agriculture, and pesticides. Before this, India wasn’t self sufficient. Introducing these lead to amazing yields on paper, India became self-sufficient on the agricultural front. Our exports increased, our economy benefitted due to self sufficiency. At that time, I don’t think we could have made a better decision. It was the need of the hour due to competitive pressure, over reliance on other nations and to protect our sovereignty. But the current time, we extract the highest amount of groundwater in the world, accounting for nearly 25% of global extraction (More than combined US and China’s groundwater extraction). If we continue like this, in a few decades we will be seeing the Indian economy in a deep state of crisis where it would be hard for us to comeback. We are already seeing food inflation. Our exploitation of groundwater has led to per capita water decreasing by half since the green revolution. It is silently and invisibly creating a crisis that we aren’t prepared to handle.

This is what I fear will happen with AI but at much faster speed than the green revolution. Today, I work in the government. We are already a very fragile public system. It’s neither perfect nor efficient, there’s huge public outcry to improve them. That’s why the central government is pushing for deployment of AI in the public systems and for AI sovereignty. The deployment may be as we did it during the DPI but it is the governance that I am most concerned about. Aadhaar’s deployment was well executed but even after 15 years we are still figuring out how to govern the Aadhaar data. We talk a lot about how this will improve efficiency, but our institutions, similar to when we adopted the green revolution, aren’t equipped for it in the long run. Farmers were pulled into a model that rewarded short-term output while hiding long-term ecological costs and they saw the boom in their revenue at the beginning. The same is happening with government officials, they will instead become verification tools while their output will increase. Slowly, the only judgement they will be making will be to tick a checkbox. This is where, the institutional failure begins similar to how farmers keep committing suicides and our nation can’t do anything to protect them. There will not be suicides but small accumulative x-risks will carry forward leading to erosion of institutions as described by Kasirzadeh.

It will begin by making officers verifiers rather than decision makers. The system will reward increased output and officers who question will either succumb to competitive pressure or will look slower. Eventually, judgement will be a liability and reasoning will move to the model.

Then the failures will accumulate quietly. Citizens will be wrongly excluded from benefits. Risk scores will misclassify people. Grievance systems will become harder to challenge because no one can explain the decision. Departments will respond by adding more checklists and dashboards, not by rebuilding judgment. The institution will still look functional, but its cognitive capacity will hollow out.

The crisis comes later, when advanced AI systems require real governance. By then, the officials meant to regulate, audit, and override these systems may have spent merely approving their outputs. The state will have efficiency on paper, but no institutional muscle left to govern the machinery it has adopted.

With the influx of new funding, don’t hunt for new problems—stop suffering first by Yaqi Grover

Yaqi Grover — Wed, 20 May 2026 07:17:25 +0000

A new article went viral on Twitter today: Nan Ransohoff’s “The Third Wave of American Philanthropy” (link). Worth reading first.

Nan is right about the shape of what’s coming: hundreds of billions in new philanthropic capital, no ecosystem yet to absorb it, and a shortage of builders and organizations. I very much agree with that sentiment and the direction. More money, more people willing to start things, more urgency.

But the conclusion I draw is a little different. The new philanthropic wave shouldn’t go hunting for problems in completely new places: it should look harder within one it already named and set aside, within an all-encompassing field that history reduced to a single term too small to carry its gravity.

Far more of the new philanthropic wave should go to helping animals. It should go to the trillions of lives in all their variety and multitude. “Animal welfare” isn’t one solved issue to cross off: it’s where most of the sentience and suffering is. It covers the vast majority of moral patients, all of them suffering gravely in every corner of the world. The future is still incredibly grim; AI could impose even more suffering if we don’t get this right.

Animals are not one issue

EAs recognized the importance of animal suffering – factory farming, wild animal suffering – long before the rest of the world, which still has not really recognized its importance. That was the insight. And then, somehow, we just decided to group it all as one single issue. One moral cause area.

Chickens, fish, pigs, shrimp, insects… farmed and wild. Nonhuman animals are the most populous category of moral patients by orders of magnitude. Yet we have collapsed all of it into a single line item that competes for attention with everything else as if it were just one intervention among many.

The problems are staring right at us

This could be the path we go down in the name of flourishing: we simply lower the bar on cost-effectiveness. We go looking everywhere for things to do, finding problems to solve, manufacturing causes to fund. But are those really the best use of a once-in-a-generation wave of capital and talent?

If we still take scale and cost-effectiveness seriously, we have so, so many problems staring right at us. They just happen not to be happening to humans.

What we need is a lot more people working on:

getting billions of chickens out of cages;
getting tens of millions of mother pigs out of gestation crates;
giving billions of ducks access to water;
stopping hundreds of millions of frogs from being dismembered alive;
preventing tens of billions of chickens from fracturing under their own weight;
sparing hundreds of millions of rabbits whose feet are sliced raw by the wire cages;
stopping trillions of feeder fish from being eaten alive;
preventing hundreds of millions of mice and rats from a slow death suffocating on their own blood;
preventing trillions of insects from being farmed in horrendous conditions (even here, I’m reducing the most numerous animal species on Earth into a single word).

Just some examples, and the list goes on and on and on. I’m listing these only to make them slightly more granular.

There are a lot more neglected animal populations that no one is doing anything about, and we should be looking for those problems (I left my job recently to do exactly this). A lot of new effort should go toward finding new ways to help all those neglected beings.

AI x Animals

Contrary to popular belief, many “animal people” in EA recognize just as much that the most critical work today lies in making sure AI goes well. But while it is easy to appreciate AI’s grave implications, what is even more alarming is something the discourse rarely touches: AI’s implications for animals might still be graver.

We could flourish while the moral atrocity that is factory farming is left intact, or even expanded by AI. Trillions of sentient lives could continue suffering completely unheeded as they stay almost entirely outside the moral circle.

Yet anything in that AI × animals intersection gets filed, yet again, under “animals.”

“AI x animals” is a sub-bucket of a sub-bucket, rather than treated as central to the long-term future we keep saying we care about. If we are serious about a future with less suffering, then what AI entrenches or prevents for animals is not a footnote to that future. It may be most of it.

Stop looking past them

Under any reasonable moral weight model, animal consciousness and suffering far trump that of humans—today and very possibly long into the future.

Beyond making sure we don’t all die, stopping animal suffering should be our first priority. We should address this before moving on to civics, art, and flourishing.

The problems are staring right at us, and we continue to look past them. And now, with all the new energy pointed at finding new problems to solve, we are about to look past them harder, with more money and more talent than ever. Let’s not.

Request for Proposals: Effective Giving by Melanie Basnak🔸

Melanie Basnak🔸 — Tue, 19 May 2026 17:28:34 +0000

We are excited to announce a new Request for Proposals (RFP) for effective giving organizations, defined as initiatives that raise funds for highly effective charities. Through this RFP, we aim to identify and support additional efforts in the effective giving space and streamline the application process for potential grantees.

Motivation

Our Effective Giving & Careers Fund has historically supported a variety of effective giving organizations across multiple countries. Examples of current grantees include Giving What We Can, Effektiv Spenden, and Doneer Effectief (you can find other recent grants here).

Effective giving currently comprises ~70% of our portfolio, reflecting its central role in our broader funding strategy. In the past few years, we have been surprised by the growth of the effective giving ecosystem: new initiatives have launched worldwide, and existing efforts have successfully channeled substantial funds to promising opportunities.

Based on our internal analysis, we estimate that our current grantees deliver an average adjusted return on donations of 6x across our effective giving portfolio, with some grantees as high as ~10x. We believe that additional organizations we’re not currently supporting may be poised to generate similar impact. Through this RFP, we hope to:

Invite and encourage new and existing effective giving organizations to apply for funding.
Simplify and standardize the application process for both applicants and our internal review team.
Better understand the current funding needs and gaps within the effective giving space.

Eligibility

Any organization contributing to raising funds for effective charities can apply. This includes, but is not limited to:

Donation platforms (and associated outreach efforts) in their respective operating countries.
Organizations advising (U)HNW donors on charitable giving.
Organizations recruiting pledgers (i.e. those encouraging individuals to pledge a percentage of their income to charity).
Organizations using matching or multiplier schemes to encourage donations to more effective charities.
Groups raising awareness of effective giving or funneling new donors to other effective giving organizations.

Additional considerations

Our current grantees raise funds across a variety of focus areas that align with Coefficient Giving’s strategic priorities (global health and development, farm animal welfare, global catastrophic risks). Any organization raising funds for those areas is welcome to apply.

Your organization should be able to justify how it can cost-effectively use additional funding.

Existing grantees on renewable grants will continue to be assessed for general support on their regular schedule. However, they are welcome to apply for top-up funding if they believe they can utilize additional funds cost-effectively. Top-up funds awarded through this RFP will be a one-time supplement rather than a permanent increase to future grant renewals.

Funding amount and grant structure

In general, we think unrestricted funding for general operating support can be the most valuable funding type, so we expect to provide that by default where possible. That said, we are open to funding specific projects or needs; also, other factors such as the grantee’s corporate structure may influence how we structure our grant.

We aim to represent at most 50% of operational funding for more established organizations, but are comfortable contributing to a larger percentage of total funding for newer organizations. Please keep that in mind when you apply for funding.

We expect most grants to be for one or two years and non-renewable by default.

We don’t have a predetermined minimum or maximum number of applications we intend to fund. We are keen to support as many promising opportunities as possible.

Ideas we’d be excited about:

New country-level organizations in promising markets (e.g. higher-income Asian countries)
Professionalized marketing support for effective giving organizations
Growth and donor advisory capacity, especially for high-net-worth donors
More ambitious partnerships with mainstream philanthropy
Targeted outreach to high-influence audiences (e.g. celebrities, public figures)
Working with corporate social responsibility arms of companies to direct funds to effective charities
Partnering with donor-advised fund platforms to nudge donors toward cost-effective giving at the point of allocation
Workplace giving campaigns and payroll giving integrations to reach donors earlier in their giving journey

Selection criteria and evaluation process

We expect to consider the following factors in our assessment:

Track record (if applicable). This will include, but not be limited to, our estimated return on donations
Strength of the theory of change and plans moving forward
Scalability and growth potential
Geographic or strategic fit with our current portfolio

Our application was designed to be comprehensive and provide all of the information we think we will need to assess the candidates. That being said, we may reach out to certain applicants for additional information.

Application logistics

You can apply using this link. If useful, you can refer to this sample application using a mix of mock and real data. Applications are due by June 26, 2026, at 11:59 pm PT.

We aim to notify all applicants of their status by the end of August. Due to the anticipated high number of applications, we won’t be able to provide individualized feedback to organizations we decide not to fund.

If you have any questions related to this RFP, please feel free to contact Melanie Basnak at melanie.basnak@coefficientgiving.org or Kearney Capuano at kearney.capuano@coefficientgiving.org. Please note that we will not be fielding questions after June 15.

CHW programs are the platform that makes EA’s bets scale by Madeleine Ballard

Madeleine Ballard — Tue, 19 May 2026 15:30:46 +0000

[N.b. I used Claude to help format this post; all arguments are mine]

A response to Rethink Priorities HSS report & F/U, with some inspiration from @NickLaing, @Berke and others.

There is a category error sitting at the centre of the EA Forum conversation on community health workers (CHWs); workers who bring care straight to their neighbor’s doorsteps. Rethink Priorities’ HSS report models “CHW programs” as an intervention, scores them at roughly 1,000x in Coefficient Giving terms, and notes they do not meet the 2,000x bar. The conclusion that follows is that CHWs are an interesting but second-tier bet.

But CHWs are not an intervention. CHW programs are the delivery platform through which most of the interventions EA already funds—bed nets, vitamin A, deworming—reach billions.

Modelling the platform against its own throughput suggests the model for the throughput wasn’t built to deliver.

A platform, not an intervention

A community health worker is not an intervention any more than a nurse is. Whether a CHW program is cost-effective depends, like any other delivery mechanism, on a) what it delivers and b) how it is designed.

What the platform delivers

A modern, professional CHW program is a primary care delivery point, not a single-disease vertical. The WHO competency-based curriculum—the closest thing the field has to a current job description–sets out the portfolio: childhood illness (diarrhoea, pneumonia, malaria), maternal and newborn health, family planning, nutrition screening, immunisation support, TB and HIV case-finding, NCD screening (hypertension, diabetes), mental health first response, neglected tropical disease case management, and health promotion.

The same worker, in the same household visit, delivers whatever subset the program is designed for.

Any cost-effectiveness analysis that attributes the full fixed cost of that worker to one of those service lines is going to produce an incomplete number.^[1]

Facility-based care alone cannot reach everyone. Hospitals struggle to serve remote communities, and in most high-burden countries the realistic counterfactual to a CHW is not a clinic visit — it is no contact with the health system at all.

Where researchers have compared CHW-delivered care head-to-head with facility-based care, the CHW route has been consistently cheaper. Most CHW programs evaluated for HIV, TB, malaria and for reproductive, maternal, newborn and child health were cost-effective against the facility comparator.

The median annual cost to deliver primary care via a CHW platform—across 380 scenarios in 130 studies—is US$0.59 per capita. That is not cost per DALY, and not cost per visit. It is the full annual unit cost, per person per year.

The median cost per beneficiary is $10.03, but of course the cost varies by service package. Integrated horizontal platforms—many service lines through one worker—sit at a median of just $6.02 per capita per year to reach those not close to a facility. ^[2]

HIV, TB, malaria and RMNCH range more widely; in both areas, CHW programs were cost-effective in over 80% of scenarios assessed.

Bundling is where the savings compound. When Liberia folded NTD case management into an existing CHW platform, cost per diagnosis fell up to tenfold versus running NTDs as a standalone vertical.

The mechanism is platform economics. A CHW’s fixed costs—salary, supervision, supplies, transport, data—exist whether she delivers one service line or ten. Vertical analyses load the whole stack onto a single disease; horizontal analyses spread it across everything the worker delivers in one visit.

How the platform is designed

Pool salaried, supervised, well-supplied, skilled CHWs together with volunteer, unpaid, under-trained ones, and the average tells you little. The “critical components” literature—including the work by yours truly that Rethink cites—has been clear for a decade: compensation, training, supervision, supply-chain integration, and career progression are not optional. Averaging two different things: poorly supported volunteer initiatives and high-performing professional platforms isn’t consistent with the documented 4x variation in ROI.

Contrary to what Nick Laing’s piece from April ’25 implies, this isn’t a workforce that is “often not cost-effective.” Like nurse-led clinics—which Nick rightly champions!—it is a platform that is cost-effective when designed properly and not cost-effective when it isn’t.

The relevant question is not whether to fund “CHWs” but which design pattern to fund, and at what level of the system.

Rethink’s way forward

For all the buzz about CHWs, the design pattern with the strongest evidence is the most neglected. A professional CHW platform is salaried (not volunteer), skilled (competency-based training to a published curriculum), supervised (relevant supervisor-to-CHW ratio), supplied (the commodities the job requires), and built into the public primary care system.

So the smart bet is the one Rethink itself names in its own report: “Help fund and establish a public health program until it is transitioned to national government ownership.”

For the whole vertical portfolio EA has spent fifteen years building—bed nets, SMC, vitamin A, deworming, child immunisation—that pathway is the scale opportunity.

How countries can get there

How does a country move from volunteer CHWs to a paid, professional, government-owned workforce? Three conditions hold the problem in place, and an EA funder can act on each one.

Guidelines. Until recently, a health minister had no authoritative international consensus describing a well-designed CHW program That has changed. The WHO 2018 guideline, the aforementioned curriculum, and proCHW guidance co-authored by Africa CDC and other norm-setters now give a minister both the cover to make a major workforce decision and the spec to make it well.

Start up capital . Even a willing government faces a chicken-and-egg problem: standing up a salaried workforce needs startup capital before the recurrent domestic budget line can absorb it. The Global Fund and others move money at that scale, often catalyzed by private actors.

Domestic constituency. Professional CHW associations are what turn a presidential commitment into a recurrent budget line that outlasts the president who made it.

Funding a delivery NGO to ship commodities saves lives now, and it is essential. Funding the workforce, the policy uptake and the government adoption that let a country deliver those commodities itself is the complementary lever that outlives any single donor.

This is a bet not dissimilar to hits-based giving: a large, durable return (i.e. a permanent, domestically-financed platform that scales all the verticals and outlives any single donor) with outcomes hard to attribute to any one funder.

It also has the one thing a hit almost never does: a decade of cost-effectiveness evidence already behind the intervention. The only open risk is whether a government finishes the handover, but if it does it pays for the whole portfolio.^[3]

A closing thought

Berke framed the underlying choice well on the Forum:

“Does success for EA look like hundreds of cost-effective NGOs distributing health commodities across South Asia and Sub-Saharan Africa until 2100 — or governments able to keep their people from dying of preventable conditions without depending on funders far away?”

For any funder serious about scale, it’s the second. And professional CHW platforms are the pathway.

^
Rethink’s CHW model, for instance, uses integrated child-survival RRs at full credit, but loads the fixed worker cost against that one outcome category. It’s likely that that CHW is also delivering services affecting, e.g., adult morbidity, non-fatal iCCM DALYs, contraception, NCD case-finding, mental health. Platforms should likely use multi-outcome accounting to reflect these real-world conditions.
^
To my point in this article: the $0.59,$10.03, $6.02 figures sit on quite a bit of heterogeneity! They are medians across 130 studies, with wide variation in setting, package, and study quality. I think the central tendency is robust enough to act on. I don’t think any single number should be quoted without that caveat, and I’d welcome a re-analysis that weights by study quality.
^
Most EA models are static; handover, adoption, and durability are not modelled. In the HSS piece, 997x is the marginal-DALY price under a permanent(?)-donor assumption. The catalytic case is the temporary-donor price: donor cost goes to zero after handover, government carries the platform for decades, and the accounting compounds across every category it produces (see footnote 1)

Nobody knows what ‘AI exposure’ means by Deena Mousa

Deena Mousa — Tue, 19 May 2026 13:09:29 +0000

I worry that the usage data on which AI exposure indices are built is sufficiently removed from labor market disruption, and what these indices actually are capturing sufficiently ambiguous to 90%+ of people, that they are actively harmful for getting policy right. I wrote something about it.

Why are longtermists so much less focused on human extinction these days? by Michelle_Hutchinson

Michelle_Hutchinson — Tue, 19 May 2026 12:31:23 +0000

Tldr;

(Longtermist) EAs seem more focused on non-extinction risks from AI than they used to be
I haven’t seen much discussion of why, making reference to the very longrun
Changes in the world this might be responding to are: alignment seeming more promising than it used to, the speed of AI progress, and concerns around US governance.
There are also plausible sociological reasons for the shift, so I feel unsure how far to update

What’s the change and why care about it?

My sense is that there’s been a significant shift in how much longtermists prioritise non-extinction risks over the last few years. A decade ago people who were trying to ensure the flourishing of the universe trillions of years from now were very often focused on avoiding events that would kill all humans. My sense is that now they’re much more often focused on situations which they don’t think pose a risk of human extinction, such as extreme power concentration.

This change has felt confusing to me for a couple of reasons:

There’s reasonably little written about why longtermists should change their prioritisation in this direction. The notable exception is this paper by Will MacAskill.
To my mind, there’s a clear reason for longtermists to focus on extinction risks:
- It seems extremely hard for any of our actions now to predictably affect what the world looks like in trillions of years time. (Events which predictably affect the very longrun future are often called ‘lock-in’ events.)
- Preventing humans from going extinct in the next decade continues to affect the future indefinitely—human extinction seems like a clear ‘lock-in’ event.

You might think that our actual actions won’t be affected much by whether the risks we should prioritise most are totalitarianism or human extinction. There are indeed many actions which are useful for both of these. Pausing AI development will give us more time to allow the world to prepare itself for all sorts of risks that transformative AI might bring. Frontier AI companies having good whistleblower provisions gives us a better hope of finding out about many of the largest of those risks before they come about.

But there are ways in which you could disproportionately affect extinction risks compared to other risks. One is working directly to mitigate the ways humans could all plausibly be killed, such as AI enabled biorisks. Another is to focus more on the risk of misaligned AI risk than the risk of AI misuse by humans. The reason is that AI may be inherently alien to humans and therefore likely to have goals which are far reaching and extremely different to ours, and therefore cause a future which doesn’t include us. “The AI does not hate you, nor does it love you, but you are made out of atoms which it can use for something else”. By comparison, humans tend to have goals for which it’s useful to have other humans surviving.

I wanted to get a sense of why there’s been a shift in favour of longtermists prioritising extreme power concentration amongst humans compared to AI takeover, in order to understand better what’s highest priority for me to focus on. Below are the reasons I came across when talking to people about what has caused this trend.

Differences in the world that motivate the change

Alignment is more promising than some people expected.

For many longtermists, by far the most likely cause of human extinction is a misaligned AI. The reason is that humanity is actually reasonably resilient to large shocks, even ones that kill the vast majority of the population. That means natural disasters are reasonably unlikely to wipe everyone out. Whereas a misaligned AI for whom humans were in the way might take a more intentional approach to wiping us out.

Ten years ago, it seemed plausible that AI wouldn’t be able to understand human goals at all, so the idea of AI pursuing totally alien goals seemed more likely. We now have empirical evidence of AI understanding humans reasonably well—being able to get a fairly brief and vague prompt and accurately determine what we’d like it to do and how. To the extent their goals diverge from ours, they’re often very intelligible do us (such as wanting to complete the task and therefore cheating on it). There are some structural reasons for that—using large language models extensively imports a lot of human concepts into models.

Even aside from alignment being less challenging than some expected, the fact that there are now more people working on it provides some reason for thinking that marginal work on it might be less useful than it was in the past.

AI progress has been continuous, and medium-fast

In a world where AI progressed from largely useless to superintelligent in the course of days, there’s little opportunity for humans to figure out how to control it or wield it to their ends. In a world where AI progress is very slow, we have time to adjust our governance mechanisms to the existence of ASI. But where progress is continuous and medium-fast, some humans have enough time to notice it would likely be useful to control super-intelligence and make plans to do so, without there being time for things like governments

US governance is in a worse state than we might have expected for the point at which we hit transformative AI.

The federal government seems unusually opposed to safety-focused AI legislation (making it more likely than it might have been that I company execs could amass unusual levels of power), and less respectful of separation of powers. It’s possible that the people running AI companies also seem more power-seeking / less safety conscious than we would have expected (though that seems less compelling to me).

How much of an update should we make?

The above all seem like solid reasons to shift some prioritisation from avoiding AI takeover to avoiding extreme human power concentration.

I’m still not sure if I buy that these risks are in the same ballpark, even if they’re now closer. My hesitation stems from a strong intuition that it’s extremely hard to predictably affect anything over the long-term. Our actions have a very strong tendency to ‘wash-out’ - often over mere years, but particularly after centuries, millennia and onwards.

There are reasons to think that AGI will make it less likely that actions wash out—for example, it will plausibly enable immortality (perhaps by uploading rather than biologically). Dictatorships have often historically been brought down by the natural death of the dictator, so the impossibility of that could really increase the chance of lock in over the very long run. (For more on how AGI might enable authoritarian lock in, see the MacAskill and Finnveden et al papers linked about.) On the other hand, technological breakthroughs have historically swung the balances of power and otherwise washed out previous actions, and transformative AI will precipitate a time of unusually rapid technological progress.

There are also various things which might have causally led to a shift in views without epistemically justifying them:

When OpenAI and DeepMind were set up, their comms about their aims were heavily focused on keeping the world safe. Things like the OpenAI board turning over has made salient how much some of the key players seem squarely focused on amassing power, and how scary that is. Being viscerally aware of that (even if the extent to which it’s the case is within the bounds of what you would have theoretically expected) makes it harder to look away from risks from human power concentration.
The US has become more divided and tribal, making it feel more urgent and important to work on preserving democracy
As AI progress is discussed more in the mainstream, there’s a larger coalition working on ensuring that the transition to superintelligent AI proceeds safely, many of whose values don’t extend to the extremely longrun future. Preventing extreme concentration of power is a unifying issue in the sense that lots of people find it more plausible than extinction, so this provides a (perhaps unconscious) incentive to emphasise it more.

The existence of causal reasons like those above make me wary of over updating in towards prioritising extreme power concentration. On the other hand, I think it’s plausible that I previously focused too hard on human extinction compared to other plausible lock-in events. A couple of the reasons I might have done that:

I have a strong intuition about it being very hard to affect things over the longrun. And there is plenty of evidence about people trying to cause lasting change and totally failing. But I think it might be a mistake to trust my intuitions about this question at all. We should expect the world to be unrecognisable after a transition to ASI, so why would we expect it it to be similarly easy/hard to lock things in for the future?
It does seem right to me that there is a chance of us being able to lock particular values or governance structures for the extremely long run, and also a chance that preventing human extinction does not lock in a change in value of the universe (for example, if humans went extinct decades later it would wash out the effect of saving them this decade). I think it’s easy to round the former off to ‘basically not lock in’ and the latter to ‘basically lock in’ in a way I don’t endorse.

All this mostly leaves me still feeling pretty unclear about how to prioritise between risks that might and might not plausibly cause human extinction. I’m inclined to prioritise those that won’t somewhat higher than I have historically. I still feel nervous that I’m partly doing that for bad reasons. Those reasons are less the explicit causal ones I listed, and more deferring to the ‘general zeitgeist’ without fully understanding its reasoning. I’d be very keen to hear other people’s views on this question.

Thanks to Arden Koehler for making this post (and so many of the things I write) much better, and to various people for the ideas behind the post, including Alex Lawsen and Nick Beckstead.

Applications open: CRS Intro to Suffering-Focused Ethics Course (6-week online program, July 13 – Aug 23) by Zoé Roy-Stang (CRS)

Zoé Roy-Stang (CRS) — Tue, 19 May 2026 11:00:32 +0000

I used AI (Claude) to help draft and structure this announcement based on our course materials.

The Center for Reducing Suffering is launching the first edition of our Intro to Suffering-Focused Ethics (SFE) Course.

What this is

A 6-week online program introducing the core ideas, arguments, and practical implications of suffering-focused ethics — the family of moral views that prioritizes reducing suffering, especially intense suffering, over creating additional happiness.

The course is suitable both for those just exploring SFE and curious to learn more, and for those who have already engaged with these ideas. It aims to build a stronger community of people thinking carefully about effective suffering reduction.

Logistics

Format: Weekly 90-minute facilitated discussion sessions in small cohorts, plus 2-3 hours of weekly reading
Dates: July 13 – August 23, 2026, with an end-of-course event the week of August 24
Cost: Free
Optional weekly Q&A sessions with experts
Optional final project (memo, blog post, video, paper) with facilitator feedback

Apply

Deadlines:

Participants: June 7, 2026
Facilitators (small stipend; mostly confirmed but a few spots may open): May 31, 2026

Questions? Contact us at courses@reducingsuffering.org

Coordinal: A Postmortem. by Ronak Mehta

Ronak Mehta — Mon, 18 May 2026 20:43:33 +0000

cart;horse: Most startups fail, this one did too. Tried too much with too little, with a little too much ego.

Coordinal Research’s goal was to build an automated safety research platform. A researcher writes “Replicate X result from paper Y with tweak Z” and the system provisions a sandboxed compute environment, gathers context, writes code, runs experiments, and returns a research report plus oversight trail.

The bet was that AI-accelerated research is coming whether we like it or not, and having this differentially happen for safety work faster is good. The bet didn’t pay off — at least not in the form I tried to make it.

The main content is the condensed version. The full chronology, the final push, and the technical artifacts I shipped are appendices for anyone who wants them. See also our Manifund page for the original pitch/ask.

The arc, in brief.

The startup. Fall 2024 to Fall 2025: out of MATS 6.0, through Catalyze, into cofounding with Jacques. Most of Q1–Q2 2025 was grant applications and demo building (no dice on grants). In April a funder I’d previously pitched came back via a coworking connection. After a carefully-crafted email, we closed $125K on an MFN SAFE. Incorporated Delaware C-corp, put on some MATS workshops, built more, joined the ⁵⁰⁄₅₀ accelerator. Split with Jacques in October, started with Leo in November, split with Leo in late January. Two splits and a lot of admin between them.

The startdown. Q1 2026: a final push on two things. (1) Ship the user-facing app at coordinal.org/app. (2) Demonstrate SOTA on RE-Bench. The RE-Bench work went well — full-suite normalized average from 0.547 to 1.624 over a month, ~$30k of compute, ⁶⁄₇ tasks reliably producing real non-reward-hacked results with Sonnet 4. The app demo didn’t — a friend couldn’t figure out the interface, which clarified I was much further from a shareable product than I’d thought. Burned out, I decided to stop unless funded. Coefficient Giving eventually declined the $1M budget I’d sent.

The salvage. A working full-stack platform on AWS, an agent orchestration system with real observability, an RE-Bench eval pipeline, and a tour through the entire legal/admin layer of incorporating and unwinding a small tech org. Detailed writeup in the appendix. Could have been worse.

Lessons.

Counterfactual safety from outside is structurally hard. Most platform/tooling-shaped safety work gets built by default by the frontier labs and the big agent frameworks, faster than an outside org can ship equivalents. An early scaffold I had was eaten by Claude Code; my tmux+worktree manager (geewit) is now essentially a CC feature; various hooks and oversight tools have since shown up in agent frameworks and provider tooling. For-profit AI safety has its own issues: as everyone in the ⁵⁰⁄₅₀ AI safety stream stumbled into, most fundable for-profit safety work is security middleware. Non-profits have a complementary problem: there’s effectively one big funder, and your job is to convince them. The broader worry is structural: selection pressure across the ecosystem — toward research credentials and known/agreed-on bets on the non-profit side, toward middleware and product on the for-profit side — systematically under-produces the kind of counterfactual work I think the field most needs.

Forever fundraising, and never salient enough to stop. $125K felt like a lot. It isn’t: more than half goes to compute and contracting, and you can’t pay two Bay Area salaries from the rest for more than six months. Six-month grant cycles with no in-cycle contact while burning $5k/mo of personal runway is hostile. The maybe interesting asymmetric lesson: small grants below the threshold of the work may be net-negative: they keep you alive enough to keep trying without enough to actually do the thing. I was too frugal early, I should have spent the money and built out the RE-Bench experiments 6-10 months earlier and ignored platform-building. I figured out too late that for the work to succeed I need proper support and to not be thinking about fundraising for at least 6-12 months at a stretch. The ecosystem isn’t really set up to provide that, and I think it should be.

I should have played the game more. I thought BOTECs for alignment work were dumb — how would estimating differential x-risk reduction basis points mean anything? I figured smart funders would just get the obvious argument around reducing elicitation overhang. They didn’t, and ego stopped me from making up the numbers anyway. Probably should have.

I tried to do too much. Org-builder + research engineer + senior cloud/infra engineer is three jobs. Counterfactual cost was close to $1M/yr at industry rates; I was doing it for free, often alone, while fundraising. The conviction-vs-flexibility tradeoff cut against me too: I cared about the thing getting built, not the org, so I couldn’t easily pivot when feedback pushed that way. Maybe no org at all would have been better — independent funding, mild affiliation, much less admin. I should have set sharper milestone-shaped goals and ignored more sidequests (MATS workshops, oversight tools, etc.).

I built for myself, and didn’t show anyone. Standard mistake; doesn’t make it less true. My internal bar is so high that I rarely thought anything was ready to show, so scope creep + 100 self-generated P0 tickets meant I sat on the work. I should have hill-climbed on concrete, externally-legible outputs (e.g., the RE-Bench numbers I generated at the end) six to ten months earlier — way easier salience, fast feedback loops, the kind of thing fundraising actually responds to.

Cofounder fit: trust your gut. I ignored mine. “You need a cofounder” is partly a hedge for funders and VCs. On the founder side, given how often startups fail anyway, you spend a lot of time with cofounders who won’t work out. Both my splits were the right call; I should have made them faster, or trusted my “full body yes” calibration earlier and gone solo from the start. (I’m genuinely lucky both splits were with great people who operated in good faith — community norms made this a lot easier than it could have been.)

What’s Next.

So I’m winding down for now. The org will be in some hibernative state (I still may want to do something in the future, the AWS credits are still useful, the codebase could be returned to). The SAFE remains in place and if we reactivate the terms still hold. The METR thread eventually faded. We talked about collaboration, they pointed me at open roles, I applied and didn’t make it through the work trial. I’m doing some contract work building uplift and automation tooling, and am thinking about where and what is the best counterfactual use of my time and skills. It’s hard for me to not just full-send on what I think is the most important/ambitious/long-term thing, and right now I’m leaning towards automated philosophy or advocacy, so we’ll see where I end up…

I think I’m bearish on starting organizations right now given the funding landscape and what I think is necessary for an org to succeed and have a counterfactual impact, but I’d be happy to talk to anyone about how their situation may be different! Feel free to reach out for a chat about any of this. I’m particularly interested in if any of these things could point toward ecosystem-level improvements to make it more likely that others in similar positions succeed, or at least don’t waste too much of their precious time.

I used an LLM to help draft this post and it likely contains >10% AI-generated text, but I’ve edited/rewritten it extensively and endorse it.
This is also on substack: https://ronakrm.substack.com/p/coordinal-a-postmortem

Appendix.

A: The Startup.

I had just finished MATS 6.0 in the Fall of 2024, was mildly working on some follow-up work, and decided to quit my MLE job that wasn’t really aligned with directly addressing the technical problems. There had of course been talk about needing more startups and founders, and Catalyze was starting so I applied. As part of the application, I built a weighted factor model. It was rough, but pointed me in a direction I hadn’t really considered, and realized was severely neglected and tractable (in my eyes): building out capacity for research automation differentially pointed at safety sooner rather than later. Catalyze was great to slowly figure out what it would look like to build an org. There was an emphasis on cofounder matching, I got to meet awesome people, learn from them, and start working with someone. We ended up splitting towards the end of January, and that’s when Jacques and I began working together.

For most of Q1 and Q2 2025 Jacques and I spent a lot of time figuring out what we wanted to do and how, splitting duties roughly along CEO/CTO lines, respectively. We spent a lot of time on grant applications. They say not to spend more than X hours, that is not reality. We worked on pitches for ARIA, OpenPhil’s TAIS (Technical AI Safety) RFP, SFF (Survival and Flourishing Fund), and Foresight. During this time I was also building out an MVP whose scope kept growing. People at EAG (Effective Altruism Global) Bay Area 2025 were interested and generally excited, but it wasn’t at a point to ‘Wow’. A funder who saw the demo saw some potential but suggested that I focus on making the interface much better (it was essentially a CLI with a rough pure JS ‘viewer’). Started spending time on that, as well as figuring out how to create an async demo I could share remotely (read: I started building a web app and figuring out what the cloud infrastructure would look like…). To be clear, this was a platform for autonomous safety research. As you might expect, this requires live compute for a ‘project’, this isn’t just a PostgreSQL database with a Next.js frontend. I needed to be able to allocate a full compute stack to run what was, at the time, a prototype agent framework with full read/write/execute permissions, custom tooling, internet access, filesystem management, and GPU access. You can see the issue here…

This was all burning personal runway, using my own money to spend on compute, and before Claude Code had been released or was any better than the scaffold I had built custom. In April, a co-working space connection reintroduced me to a funder who had originally passed during my Catalyze pitch (Jacques’ as well). The pitches weren’t great, and I understood that decision. Fortunately, with a carefully crafted email (spent about 3 hours on this one email), I was able to convince this funder that we were worthy of seed funding.

This was awesome! With some back and forth, plus some additional support from a few others, we were able to secure an initial investment of $125K. We took some time getting these funds to actually do something with; we still had massive uncertainty that wasn’t clarifying regarding whether to go non-profit or for-profit (see Jacques’s great post on this at the time). At some point I was tired of this and I just spent a day creating a Delaware C-Corp via Stripe Atlas. At the time, I was a bit frustrated with the non-profit AIS funding ecosystem, particularly around grant response times, and we were becoming a bit more concerned that scaling in the future would be difficult unless we had proper seed funding (that could potentially only come at a fast rate if we were a for-profit). Incorporated as ⁵⁰⁄₅₀ founders (I CTO, Jacques CEO), opened the bank account, received the funding via an MFN SAFE (Most Favored Nation Simple Agreement for Future Equity). This at least gave us some breathing room to pay for all of the tools we were using (at the time a lot of experimentation, close to $2k/month for each of us (see Jacques’ posts for more detail)) plus the compute costs (AWS EC2/S3/etc. alongside straight tokens).

With this money we felt some amount of comfort to build, but were frugal and conservative, because we had uncertainty about if we were going to pivot, Jacques’ ability to work as a Canadian citizen, and if we would need those funds fully for compute to push for the next fundraise. During the summer we put on some workshops on using Claude Code for MATS: I think these were generally not that successful, but it was a great learning experience.

We had decided to pursue the for-profit direction a bit more in August, building parallel MVPs (Jacques more toward uplift and Claude Code extensions, myself more toward full-stack research automation) and applying to accelerators and networking for VC intros. We were hoping to submit some results to Neel’s MATS stream application, but scope creep and failing to retrofit to the type of research needed for the work trial meant this failed. We started the ⁵⁰⁄₅₀ accelerator in September, alongside sprinting on a new product direction we felt could work very well as a for-profit startup (agent oversight products: at the time no one was using Claude’s hooks and we saw potentially high leverage in converting our internal tools to a product). ⁵⁰⁄₅₀ was great, but as we and everyone else in the AI safety ‘stream’ learned, you have to massively sacrifice on the safety vision if you want strong funding, or you have to just work on non-dilutive funding (e.g., fellowships, grants) until you have proof-of-concept and product-market fit to pivot out to venture support.

(I’m not going to characterize the cofounder splits too much, and I did share this with both of them before posting. You can assume standard reasons why such things don’t work out.)

In October Jacques and I split. Outside of the obvious reasons, this was tough — winding down a ⁵⁰⁄₅₀ incorporation cleanly takes time and care. We worked through it amicably.

I was fully set on only building the vision in my head at this point, and did not want to be distracted at all by for-profit incentives. I also met Leo around this time, and in November we started working together. I think we generally had alignment, and we spent a lot of time figuring out if we were on the same page about tons of stuff. We decided to go forward, and started talking with OpenPhil/Coefficient Giving in December. These discussions were promising but still slow.

Towards the end of January I started recognizing the same feeling I had had in October, and decided to part ways with Leo. Again this sucked, Leo’s awesome, but it just wasn’t clicking. I was also becoming highly uncertain of what was important, the original theory of change, and if I wanted to continue working on the startup. This was slightly easier admin wise, but still somewhat costly in time and money (IP agreement language, fair compensation, etc.).

B: The Startdown.

At this point I decided to make a final push on two pieces. I had been extremely distracted from what I felt was the core output that I wanted: to demonstrate automated safety research was possible. So I put my full effort into 1) building out the full end-to-end app that anyone could use, and 2) running my backend scaffold and system on RE-Bench to demonstrate SOTA performance.

So I grinded this out, hoping to be able to use the outputs of this to convince Coefficient Giving to fund me. I spent most of my time working on the frontend-backend infrastructure, debugging full-stack deployments, setting up account management, research project management, monitoring and oversight interfaces, and getting the end-to-end workflow of an end-user in a reasonable state. I also decided it was time to spend money (AWS Quota limits stopped me from taking advantage of Activate Credits, so I had to just go with Lambda Labs out of pocket) on running and hill-climbing on RE-Bench. RE-Bench results were rapidly improving as I patched holes in my interface, figured out workarounds for their Inspect integrations, identified scaffold improvements, etc. I ended up running about 17 full-suite runs over the 7 tasks, with about 77 other single-task probes. Removing outliers, my full-suite average norm went from 0.547 to 1.624 over the course of a month, spending about $30k on compute during this period. I’m also happy with the reliability improvements over that time, with ⁶⁄₇ tasks confirmed to be real results without technical issues or reward hacking every single run. The underlying model here was claude-sonnet-4-20250514, so I was fairly excited about these results.

Mid-February I hit a semi-self-imposed deadline to demo the user-facing app to a friend. It was generally working, you could copy-paste a google doc into a box and kick it off after logging in to https://coordinal.org/app/ . This would do a ton of cloud and project setup, provision a VM and container, mount any existing project files, and kick off the scaffold. The oversight was still a bit janky, but it generally worked.

They couldn’t figure out the interface, obviously. They didn’t know how to use it, there were too many settings to choose from, it was unclear what the scope of a project or run should be (even though I was intentionally trying to have it be “literally put anything in this box”), and I didn’t have a practiced pitch or tutorial ready to explain it. The demo confusion clarified I was much further from a shareable product demo than I’d thought.

I was burned out at this point and realized it was time to stop. This wasn’t working and wasn’t sustainable. I went for a walk and generally decided that I wouldn’t be working any more on this unless it was funded. It was clear I couldn’t keep doing this all alone with no money if I wanted to build this platform, and if I just wanted to do capability elicitation I should consider working with/for METR (we had started some discussions at this time). I shared the RE-Bench results above with them, but didn’t get a clear follow-up, and didn’t push — newer model cards already match or beat these, so there isn’t much point pursuing it further.

Over the next few weeks I started doing some soul-searching; the models were just getting better and any alpha I was bringing was getting more and more swallowed by Claude Code and related tools plus better models. I was pretty burned out and was not very interested in continuing working alone, so I just started looking for other existing orgs that I felt excited about, whose mission I could help/accelerate. Discussions with CG continued, they were interested in a budget for an org that would work on building out automated research replication capacity. I sent them an ambitious budget for $1M for a year, that roughly broke down to salaries for 3 engineers/researchers, ~2 FTE equiv. contractors, 0.5 FTE COO, and then $250k compute for building out PaperBench-style CI/CD workflows and infrastructure. They eventually said no (April), citing not enough research experience; I was tired enough and had already mentally moved on so let it be.

C: The Salvage.

Holy shit I learned a lot. It’s not all bad! Look at all of these things I’ve built, new skills I have, topics and ideas I can speak confidently about, ways of thinking that will contribute strongly to any future work:

Shipped a real full-stack web app with cloud infrastructure to the public internet. coordinal.org/app/ — Clerk-based auth, multi-tenant RBAC, real-time job monitoring, file browsing, goal-hierarchy views, benchmark analytics dashboards. Mostly vibe-coded (React/TypeScript/Vite/Tailwind frontend, Flask backend, WebSockets); the full codebase is a monorepo with 350k LoC (probably closer to 1M written and discarded across 18 months of agent-assisted iteration). I now actually know how to deploy something real behind CloudFront with auth, not just a static GitHub page. Warm EC2 instance pools with GPU fallback across p3/g4dn/g5 (1-5s warm starts vs. 60-90s cold-start provisioning), EFS-shared workspace storage, S3 for run artifacts and progress tracking, CloudFront + Route53 + SES + Secrets Manager. Pulumi Automation API for IaC, GitHub Actions OIDC for AWS auth (no long-lived creds anywhere), multi-architecture ECR builds. I can now estimate AWS Bedrock quotas, hit them, file the right increase requests, and route around the ones I can’t get.
Built a real agent orchestration platform with run observability. Hierarchical multi-agent system and depth-limited delegation. Four configurable continuation policies (single-shot, fixed-iteration, time-budget, goal-driven), plus custom message-history compaction with prompt-cache integration so long runs survive past the context window. Three-container Docker architecture with strict isolation between orchestrator, agent, and workspace; JJ (Jujutsu) integration for immutable time-travel snapshots at every agent decision boundary. Full OpenTelemetry/Logfire instrumentation with a hierarchical span structure (run → agent/goal → tool → checkpoint), automatic reasoning capture at every key decision point, and structured post-run reports with root-cause analysis. Every tool call, every delegation, every error is traceable.
A real RE-Bench eval pipeline, plus the results. End-to-end pipeline on top of METR’s Inspect AI framework: normalized scoring per their paper’s formula, S3-based distributed progress tracking, per-task TOML configs, multi-replicate execution, score-log sanitization so the agent can’t binary-search the ground truth. Ran ~17 full-suite passes and ~77 single-task probes. By the end, ⁶⁄₇ tasks were reliably producing real, non-reward-hacked results, with ³⁄₇ significantly above the reference implementation. Bigger win: I now know what scaffold hardening actually looks like in practice: eliminating ground-truth leakage, gracefully recovering from context-overflow API errors, locking down score-log permissions, fixing scoring-script fallthrough bugs.
The non-technical scaffolding of a tech org. Incorporating a Delaware C-corp via Stripe Atlas, Google Workspace admin, backoffice and legal vendor management, payroll, state and federal taxes, IP agreements, cofounder agreements. Annoying but useful experience, but I now know how the legal/admin layer of a tech org actually works and how to navigate two cofounder splits without destroying the relationships or the org.

Not to mention all of the mini-projects that failed but taught me something anyway. Pretty sweet, I think.

Can EA Frameworks Unlock India’s $4.2 Billion CSR Pool? by SK Saiful

SK Saiful — Mon, 18 May 2026 16:52:45 +0000

India’s ₹35,000 Crore Corporate Social Responsibility (CSR) Mandate: An Underleveraged Frontier for Effective Giving?

A Note Before You Read: I’m writing this post partly to think through a problem I’ve been sitting with for a while, and partly to find out if anyone else is working on it—or has already tried and learned hard lessons. This is also an early public draft of a working paper I’m developing. I’d genuinely love to hear both pushback and interest in collaborating.

Disclosure: I used Grok AI to assist with drafting and editing portions of this post. All core arguments, research, data sourcing, and analysis were developed, reviewed, and modified by me.

The Basic Observation

India’s Companies Act, 2013 mandates that companies meeting certain financial thresholds—a net worth greater than or equal to ₹500 crore, a turnover greater than or equal to ₹1,000 crore, or a net profit greater than or equal to ₹5 crore—must spend at least 2% of their average net profits on Corporate Social Responsibility (CSR) activities. In FY 2023–24 alone, 27,188 companies spent approximately ~$4.2 billion (₹34,908.75 crore) - an official figure drawn directly from FY 2023-24 corporate filings on the MCA National CSR Portal Data Registry, not an estimate—across 59,633 projects mapped across Schedule VII’s developmental focus disciplines, according to the official Ministry of Corporate Affairs National CSR Portal Data Registry. Since the mandate’s inception in 2014, the cumulative CSR deployment has crossed ~$26 billion (₹2.18 lakh crore) (the figure referenced in this post’s title rounds the annual FY 2023-24 number; the cumulative total is significantly larger), an aggregate baseline verified in the recent Ministry of Corporate Affairs Official PIB Union Disclosure. This marks one of the largest directed pools of corporate social investment anywhere in the world—comparable in scale to several major bilateral aid programmes.

And most of it is being allocated without anything resembling rigorous cause prioritisation or cost-effectiveness thinking.

To put a realistic number on the counterfactual gap: the explicit “Health Care” vertical alone accounts for roughly 20.5% of total annual CSR disbursements, representing ~$860 million (₹7,150.81 crore) - an official figure derived directly from the MCA’s sector-wise FY 2023-24 disclosure data, not a modelled estimate—in annual health sector spend, as mathematically verified via Annexure-II of the Ministry of Corporate Affairs Lok Sabha Statutory Reporting Registry. If we expand this definition to include adjacent statutory tracks for sanitation, safe drinking water, and malnutrition eradication, the pool climbs to ~$1.09 billion (₹9,087.42 crore). If even 10% of that baseline healthcare pool were redirected from typical visibility-driven interventions to GiveWell-tier cost-effective programmes—a conservative assumption—at standard Indian health system cost-effectiveness thresholds of ₹20,000-50,000 per DALY averted, redirecting just 10% of that pool (approximately ~$86 million / ₹715 crore) would imply between 1.43 million and 3.58 million DALYs averted annually. Even under significantly more conservative assumptions about cost-effectiveness or the share of redirectable capital, the implied impact remains in the hundreds of thousands of DALYs averted per year.

This direction is highly consistent with empirical models mapped in the J-PAL South Asia Research Database summary on Frontline Delivery Loops. This evidence base includes Abhijit Banerjee et al.’s (2010) open-access evaluation on “Improving Immunisation Coverage in Rural India” published in the BMJ, alongside Zulfiqar A. Bhutta et al.’s (2013) landmark paper, “Evidence-based interventions for improvement of maternal and child nutrition,” published open-access in The Lancet. The 3–5x DALY multiplier is an illustrative estimate; the working paper will model this formally with sensitivity bounds based on standard World Health Organization Cost-Effectiveness Alternative Thresholds. The geographic constraint is respected. The counterfactual gap is large. That delta is what constrained maximisation captures—and it is entirely invisible to a CSR committee without an analytical framework.

I currently work at Fortify Health, a GiveWell-supported nonprofit working on reducing iron-deficiency anemia in India through wheat flour fortification. I should flag upfront: I work in the public health-nutrition space that this framework would partially cover, which creates a potential conflict of interest. The analysis below aims to be cause-neutral within Schedule VII constraints, and I’ll address this directly in the working paper by testing whether the framework recommends interventions like Fortify Health’s more than the evidence independently warrants. Readers should weigh that context.

What’s Actually Going Wrong with Indian CSR?

A lot of the problem is structural rather than a failure of intent. Six things I’ve observed and read about:

1. It’s designed as a compliance mechanism, not an impact mechanism.

The legal framework under Schedule VII of the Companies Act, 2013 specifies eligible activity categories broadly—education, healthcare, rural development, environmental sustainability—but provides no guidance on prioritisation within those categories, no cost-effectiveness requirements, and no meaningful impact measurement obligation.

The 2021 amendments moved from a flexible “comply or explain” regime to a strict, penalty-backed “comply or transfer” architecture, as outlined in the official Ministry of Corporate Affairs CSR Policy Amendment Rules. Unspent CSR funds must be transferred to government-designated escrow accounts within six months of the fiscal year-end—and defaulting corporate officers face direct, mandatory personal financial liabilities. The Ministry of Corporate Affairs has issued compliance show-cause notices and initiated penal proceedings against defaulting companies; this is not an abstract threat.

The consequence is predictable and important: CSR heads operate under what amounts to a structural sword of Damocles. Speed of capital absorption and safety of the audit trail completely crowd out any optimization for impact per rupee. The rational response, from a compliance risk standpoint, is to partner with a known, reputable implementing agency and produce clear output metrics—classrooms built, patients treated, trees planted—regardless of whether those outputs translate into meaningful long-term outcomes. This is not irrationality; it is rational behavior under a badly designed incentive system.

But the same amendments also introduced something vital: a mandatory independent impact assessment requirement for large CSR spenders (companies with an annual CSR obligation greater than or equal to ₹10 crore). This requirement, tracked comprehensively within the governance frameworks of the Institute of Company Secretaries of India (ICSI) Statutory Bulletin, serves as our primary regulatory wedge.

2. CSR heads face the wrong incentives.

Even where legal latitude exists, CSR heads are rarely incentivised to maximise impact per rupee. They are incentivised to minimise compliance risk and maximise corporate visibility. An Effective Altruism-aligned approach requires embracing probabilistic thinking—acknowledging that some interventions might not work as expected but carry massive upside potential. Corporate boards, by contrast, demand audit-proof, predictable outputs: “We built 10 classrooms.” “We planted 5,000 trees.” The incentive structure actively works against rigorous prioritisation.

3. The NGO sector has a capacity problem that limits evaluability.

GiveWell’s model depends on cost-effectiveness analysis anchored in rigorous evidence—randomised controlled trials, detailed financial modelling, and tight monitoring data. When GiveWell’s own team attempted to evaluate local Indian NGOs, they documented a severe structural mismatch in their historical assessment, Evaluating Local Charities in India: standard cost-effectiveness criteria simply did not fit organizations lacking the capacity for extensive monitoring, evaluation, and documentation.

While certain metrics have improved, the fundamental measurement and evaluation (M&E) infrastructure gap remains unresolved. India has hundreds of thousands of registered civil society organizations listed on platforms like GuideStar India, but the proportion with the institutional architecture to manage large, multi-year programs and report transparently against strict outcome indicators is considerably smaller. This creates a circular bottleneck: evaluators cannot analyze what cannot be measured, and grassroots NGOs do not invest in complex measurement systems because corporate donors do not demand or fund them.

4. The intermediary layer is the actual bottleneck—and it’s largely invisible to EA analysis.

This is the structural insight I think is most underappreciated in existing discussions of this space, and the one I’d most welcome pushback on from people with direct CSR consultancy experience.

Corporate boards rarely interface with implementing organisations directly. Instead, they rely on institutional CSR consultancies and intermediary agencies—firms like Samhita, Sattva, or KPMG India’s CSR practice—to manage their allocations. Because these intermediaries are themselves evaluated on capital deployment speed and compliance safety rather than marginal outcome optimisation, they default to low-risk, asset-heavy, highly visible interventions. This reality is documented in Sattva Consulting’s India CSR Outlook Insights, which reveals that three-quarters of all corporate capital remains heavily concentrated in traditional, output-centric delivery loops. The incentive misalignment doesn’t end at the CSR committee—it runs through the entire advisory ecosystem that sits between corporate capital and ground-level impact.

The implication is uncomfortable: an EA-aligned information infrastructure that targets only the internal CSR team will not move the money. It must actively reshape the evaluation rubrics used by these consultancies—the layer that acts as the real gatekeeper to these funds. That means engaging Samhita, Sattva, and KPMG India’s CSR practice directly, not just the committees they advise. I don’t yet have a fully developed theory of change for how to do this, but the contours of one are beginning to take shape. A plausible near-term pathway has three components: first, publishing a methodological framework (the Comparative Cost-Effectiveness Registry described later in this post) that intermediaries could adopt voluntarily as a quality differentiator in a competitive market where most CSR consultancies offer nearly identical compliance-focused services; second, identifying one intermediary firm willing to pilot an EA-aligned impact assessment on a single live project, generating a case study that demonstrates the audit-defensibility benefit to other firms in the market; and third, working through the Institute of Company Secretaries of India (ICSI) and similar professional bodies whose statutory bulletins already shape how CSR consultancies interpret compliance requirements, to embed cost-effectiveness considerations into the recommended impact assessment templates these professional bodies issue. This is a sketch, not a fully tested theory of change. I would value input from anyone who has tried any version of this.

5. The evaluator landscape is emerging—but lacks the distribution to reach corporate boardrooms.

The picture is more nuanced than simply “there’s no GiveWell for India.” Platforms like GuideStar India assess governance and financial efficiency, while specialized verification bodies operating under the Credibility Alliance National Accreditation Framework focus on operational due diligence and minimum voluntary norms. Crucially, Impactful Giving India’s Evidence-Based Methodology has established a rigorous domestic effective-giving pipeline, utilizing scale-solvability-neglectedness frameworks to vet and direct funds to highly cost-effective domestic charities.

The gap, then, is not simply the absence of evaluators. It’s that existing evaluators lack the distribution networks, corporate credibility, and regulatory framing to reach mainstream CSR committees. A company under Schedule VII compliance pressure isn’t browsing EA-adjacent platforms. They need a trusted interlocutor who speaks the language of compliance, audit, and corporate governance while applying the analytical rigour of cost-effectiveness thinking. That bridge doesn’t yet exist at scale.

6. Deadline-driven allocation crowds out thoughtful prioritisation.

Many CSR teams begin their allocation process late in the financial year. The result is that partner selection happens under severe time pressure, through referrals and informal networks, with little room for systematic analysis.

The root cause is not a lack of funds; it is a structural artifact of administrative planning timelines. Under the strict penalty frameworks detailed in the Ministry of Corporate Affairs Official PIB Union Disclosure, any unspent capital not committed to an ongoing project must be transferred to a government escrow account within six months of the fiscal year-end. This creates an intense year-end administrative rush where the certainty of immediate capital absorption is prioritized over long-term strategic value. EA’s multi-year planning horizon is structurally difficult to reconcile with the rigid 12-month fiscal calendar that governs statutory CSR compliance.

What I’m Trying to Think Through

The core hypothesis—and why it’s harder than it looks.

My working hypothesis is that Effective Altruism-style prioritisation is legally viable within India’s CSR framework. Schedule VII is a floor, not a ceiling. The law dictates which broad buckets money must be spent in, but it is completely silent on how effectively those buckets are filled. A company mandated to spend ₹5 crore on healthcare in a specific district of Maharashtra can legally choose to spend it on a high-evidence, cost-effective intervention—like maternal iron supplementation or community health worker programmes—rather than an expensive, low-impact mobile medical van. That legal latitude exists.

But legal viability and institutional viability are different things. The honest framing of this project is: what would it take to close the gap between the two?

EA as a decentralised information ledger, not a moral imposition.

One reframe that I think matters enormously for how this gets adopted: EA frameworks shouldn’t be presented to CSR committees as a moral obligation or a critique of their past choices. They should be presented as a solution to a problem CSR committees already experience and already want solved—the problem of not knowing whether their money is doing anything.

CSR heads in Mumbai or Delhi are trying to allocate capital to distributed local problems-nutritional stunting in a tribal block of Jharkhand, waterborne disease in a peri-urban slum in Odisha-that they have no direct expertise in, through implementing partners they cannot independently evaluate. This is a classic Hayekian knowledge-dispersal problem: the information needed to allocate well is dispersed, local, tacit, and entirely inaccessible through normal corporate data systems.

As Friedrich A. Hayek famously demonstrated in his landmark 1945 paper, “The Use of Knowledge in Society”, originally published in The American Economic Review, central allocation of capital to local problems it cannot see clearly is structurally bound to fail due to bad epistemics rather than bad intent. To resolve these centralized information vulnerabilities within decentralized development environments, our optimization model builds on the institutional design principles laid out in Elinor Ostrom’s pioneering research on complex economic systems, accessible via the official Nobel Prize archive in her open-access text, “Beyond Markets and States: Polycentric Governance of Complex Economic Systems.” EA frameworks, in this framing, are an information infrastructure that solves a knowledge problem corporations already want solved. They don’t need to gather evidence from scratch on whether community health workers reduce under-5 mortality in low-income rural settings; that evidence already exists, has been synthesised, and can be applied locally. That reframe—institutional utility, not moral philosophy—might matter more for adoption than any argument about effective altruism as an ethical commitment.

The 2021 Impact Assessment Mandate as a wedge—turning a compliance cost into a data engine.

The 2021 CSR amendments mandated that companies with annual CSR obligations greater than or equal to ₹10 crore must conduct independent third-party impact assessments for individual projects greater than or equal to ₹1 crore. Corporate India is already paying for third-party evaluation. The problem is that most of this evaluation is uninspired compliance-checking—box-ticking on outputs rather than rigorous analysis of outcomes or cost-effectiveness. It is an administrative cost centre, not an information asset.

But it doesn’t have to be. Effective Altruism frameworks could be positioned not as an additional burden on already-stretched CSR teams, but as a way to make the mandatory impact assessments companies are already commissioning actually useful for future allocation decisions. An impact assessment designed around cost-effectiveness principles doesn’t just satisfy the regulator—it generates the comparative data that allows a CSR committee to make a better decision next year. It transforms a backward-looking compliance exercise into a forward-looking allocation tool. And crucially, a rigorous, documented, evidence-informed process is a far stronger audit defence for board directors than “we gave to a charity we’d heard of.”

This is the wedge. Not an argument about EA philosophy—an argument about making a mandatory regulatory requirement more useful, tracked directly via the compliance pathways outlined in The Institute of Company Secretaries of India (ICSI) Statutory Bulletin.

Constrained Maximisation—working within the limits rather than around them.

The approach I find most honest and most likely to be adopted is what I’d call “constrained maximisation”: accepting the geographic and sectoral constraints that Schedule VII and corporate culture impose, and then applying rigorous cost-effectiveness thinking within those constraints.

To make this concrete: a standard default intervention in this space is a high-visibility mobile medical van serving a factory catchment area-typically costing around ~$48,000 (₹40 lakh), treating roughly 1,000 patients per year, with well-documented limitations including low medication adherence, high operational costs, and minimal impact on underlying disease burden. The same ~$48,000 (₹40 lakh) directed to community health worker deployment and maternal iron-folic acid supplementation in the same catchment area plausibly generates 3–5x the DALY impact at equivalent cost.

This direction is highly consistent with empirical models mapped in the J-PAL South Asia Research Database summary on Frontline Delivery Loops, which aggregates randomized evaluations on frontline delivery frameworks. This evidence base includes Abhijit Banerjee et al.’s (2010) open-access study on “Improving Immunisation Coverage in Rural India” published in the BMJ, alongside Zulfiqar A. Bhutta et al.’s (2013) landmark paper, “Evidence-based interventions for improvement of maternal and child nutrition,” published open-access in The Lancet.

The 3–5x DALY multiplier is an illustrative estimate; the working paper will model this formally with sensitivity bounds based on standard World Health Organization Cost-Effectiveness Alternative Thresholds. The geographic constraint is respected. The counterfactual gap is large. That delta is what constrained maximisation captures-and it is entirely invisible to a CSR committee without an analytical framework.

There is a genuine tension here between EA’s cause-neutral global framing and the regulatory and cultural pressure for CSR to benefit local communities. I don’t think the right answer is simply “cause-neutrality wins.” The regulatory constraint is real and isn’t going away. Constrained maximisation is a more modest ambition than cause-neutral EA—but it’s honest, legally compliant, and might actually get used. Whether that framing concedes too much is one of the things I’d most like to hear pushback on.

Is a GiveWell-for-India feasible?

Partially, and the prerequisites matter. The NGO M&E capacity gap is real and will take years to close. But a lighter-touch version-what we might call a Comparative Cost-Effectiveness Registry for Schedule VII-is achievable in the near term.

Instead of requiring under-resourced grassroots NGOs to generate original experimental data, this entity would specialize in evidence localization: taking robust empirical evidence from the J-PAL Randomized Evaluation Evidence Repository and the Cochrane Collaboration Library, and mapping those impact coefficients onto local Indian NGO cost structures and delivery models. The output would be a sector-specific, Schedule VII-mapped comparative database-not cost-per-life-saved estimates requiring years of original RCTs, but a “cost-per-beneficiary-outcome” tier that lets a CSR committee compare interventions in their geographic mandate area by evidence base and estimated cost-effectiveness.

For evidence localization methodology, this mapping model adapts the Bayesian hierarchical tools outlined in Rachael Meager’s (2022) landmark paper, “Aggregating Distributional Treatment Effects: A Bayesian Hierarchical Analysis of the Microcredit Literature,” published in The American Economic Review, which directly addresses the econometric and statistical challenges of pooling and generalizing treatment effect heterogeneity across diverse geographic contexts.

Critically, this Registry could plug directly into the mandatory impact assessment ecosystem introduced by the 2021 amendments-generating audit-defensible, evidence-grounded comparative data that tracks smoothly against the statutory provisions broken down in the Bombay Chartered Accountant Journal’s CSR Rule Analysis and the compliance parameters detailed in the Institute of Company Secretaries of India (ICSI) Statutory Bulletin. This gives corporate boards a legitimate, legally recognized regulatory architecture to justify funding unglamorous, high-impact interventions over high-visibility, low-impact ones.

One honest caveat: evidence localization from globalized, tightly controlled trial environments to highly variable local Indian delivery contexts involves serious transportability challenges. Microeconomic parameters vary, local cost architectures differ significantly, and programmatic implementation fidelity is rarely uniform. The working paper will grapple with these external validity constraints directly by modeling structural parameter bounds against official World Health Organization Cost-Effectiveness Alternative Thresholds rather than assuming a frictionless transfer of impact.

What I’m Working On

I’m developing a working paper tentatively titled “From Compliance to Impact: Applying Effective Altruism Principles to India’s Mandatory CSR Ecosystem.” My aim is to produce something useful for both the Effective Altruism research community and practitioners—CSR committee members, corporate legal teams, and corporate foundations who might actually want to give better but do not know where to start.

The final paper will do three specific things this introductory overview does not:

Develop the DALY quantification model with proper methodology and sensitivity analysis.
Build out a concrete theory of change for how EA-aligned evaluation standards might realistically propagate through the intermediary consultancy layer.
Address the conflict-of-interest question directly by testing whether our constrained maximisation framework recommends large-scale fortification interventions more than the empirical evidence independently warrants.

I’m in the early stages, and I am not aware of anyone else currently exploring these specific questions. I’m writing this post precisely because I genuinely want to find out whether others have already worked through parts of this problem, what they found, and if there are lessons to be learned.

What I’d Love to Know from You

I’d really love to hear from you if you fit any of the following profiles:

If you’re a researcher or writer:

LMIC Institutional Giving: Are you already working on EA frameworks applied to corporate or institutional giving in low- and middle-income countries? I’d love to compare notes and avoid duplicating work.
The Epistemic State: Have you engaged seriously with the Hayekian knowledge-problem framing in an EA context? What is the current state of that literature?
Transportability Barriers: What do you see as the strongest methodological objection to evidence localisation—specifically, applying global RCT impact coefficients to local Indian delivery contexts?
Existing CEA Literature: Do you know of existing, rigorous cost-effectiveness analysis of Indian NGOs that goes beyond basic governance-and-compliance ratings?

If you’re a practitioner—in CSR, a corporate foundation, or an Indian NGO:

Operational Friction: Have you tried applying evidence-based cost-effectiveness thinking inside a live CSR programme? What worked, and what hit walls?
Reshaping the Gatekeepers: Has anyone successfully influenced or changed the evaluation and allocation rubrics used by a major institutional CSR consultancy like Samhita or Sattva? I would especially like to hear about this.
The Impact Mandate Wedge: Have you seen the 2021 impact assessment mandate used meaningfully to shift capital, or is it purely compliance theatre so far?
Boardroom Triggers: What does a corporate CSR committee actually need to see on paper to start making better allocation decisions?

If you’re a funder or EA-infrastructure person:

Applied Research Funding: Is there existing funding or interest for applied research on EA movement-building in India or the broader LMIC institutional giving space?
Infrastructure Viability: Would a Comparative Cost-Effectiveness Registry for Schedule VII be the kind of structural infrastructure project that EA Funds, Co-Impact, or similar grant pools would consider evaluating?
Strategic Bridges: Are there existing organisations currently trying to bridge the gap between EA frameworks and corporate giving that I should know about?

Thanks for reading. I’d especially welcome pushback on the framing above—particularly on whether the intermediary layer argument holds up among people with direct CSR consultancy experience, and whether constrained maximisation concedes too much to regulatory constraints that might be more malleable than I’m assuming. I’m thinking in public here and may be missing things that are obvious to people who’ve worked in this space longer.

Tagging a few organisations and individuals who may find this relevant or useful: @GiveWell, @Coefficient Giving, @Ambitious Impact, @Luke Freeman

Bibliography

Abdul Latif Jameel Poverty Action Lab (J-PAL). (2018). The Role of Cash Transfers in Improving Child Health: A Review of the Evidence. J-PAL South Asia Research Review Publications. https://www.povertyactionlab.org/sites/default/files/review-paper/CaTCH_review-paper_cash-transfers_2018.10.09.pdf
Abdul Latif Jameel Poverty Action Lab (J-PAL). (2026). Randomized Evaluation Evidence Repository Database. Massachusetts Institute of Technology. https://www.povertyactionlab.org/evaluations
Banerjee, A., Duflo, E., Glennerster, R., & Kothari, D. (2010). Improving immunisation coverage in rural India: Clustered randomised controlled evaluation of teams providing reliable services and incentives. BMJ, 340, c2220. https://www.bmj.com/content/340/bmj.c2220
Bhutta, Z. A., Das, J. K., Bahl, R., Lawn, J. E., Salam, R. A., Paul, V. K., … & Black, R. E. (2013). Evidence-based interventions for improvement of maternal and child nutrition: What can be done and at what cost? The Lancet, 382(9890), 452–477. https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(13)60996-4/fulltext
Bombay Chartered Accountant Journal (BCAJ). (2021). CSR Rules Amendment: An Analysis. BCAJ Statutory Compliance Archive. https://bcajonline.org/journal/csr-rules-amendment-an-analysis/
Cochrane Collaboration. (2026). Cochrane Library of Systematic Reviews. Wiley Online Library. https://www.cochranelibrary.com/
Credibility Alliance. (2026). National Accreditation Framework: Minimum Voluntary Norms and Operational Due Diligence Standards for Good Governance. https://credibilityalliance.org/
GiveWell. (2011). Evaluating Local Charities in India. GiveWell Official Blog Archive. https://blog.givewell.org/2011/05/04/evaluating-local-charities-in-india/
GuideStar India. (2026). Nonprofit ratings, regulatory compliance, and philanthropy transparency portal. Civil Society Information Services India. https://www.guidestarindia.org/
Hayek, F. A. (1945). The Use of Knowledge in Society. The American Economic Review, 35(4), 519–530. Free open-access text available via University of Chicago: https://home.uchicago.edu/~vlima/courses/econ200/spring01/hayek.pdf
Impactful Giving India. (2026). Our Approach: Evidence-Based Methodology and Allocation Frameworks. Impactful Giving Portals. https://www.impactfulgiving.in/approach
Marseille, E., Larson, B., Kazi, D. S., Kahn, J. G., & Rosen, S. (2015). Thresholds for the cost-effectiveness of interventions: Alternative approaches. Bulletin of the World Health Organization, 93(2), 118–124. Direct open-access PDF available via the Center for Global Development: https://www.cgdev.org/sites/default/files/Marseille%2CLarson%2CKazi%2CKahn%2CRosen-2015-Thresholds-for-the-Cost-Effectiveness-of-Interventions-Alternative-Approaches.pdf
Meager, R. (2022). Aggregating Distributional Treatment Effects: A Bayesian Hierarchical Analysis of the Microcredit Literature. The American Economic Review, 112(6), 1818–1847. Free open-access version available via London School of Economics: https://www.lse.ac.uk/economics/Assets/Documents/finance-and-development-workshop/meager-aggregating-distributiona-treatment-effects.pdf
Ministry of Corporate Affairs. (2021). Companies (Corporate Social Responsibility Policy) Amendment Rules, 2021. Government of India Gazette Mirror. https://www.sukhjitgroup.com/assets/uploads/files/pages/pdfs/88a4f-csr-mca-notification-dt-22-01-2021-companies-csr-policy-amendment-rules-2021.pdf
Ministry of Corporate Affairs. (2025). Development Sector-Wise and State-Wise Corporate Social Responsibility Disclosures: Lok Sabha Unstarred Question No. 39 Annexure-II. Parliament of India. https://sansad.in/getFile/loksabhaquestions/annex/186/AU39_PgypUR.pdf?source=pqals
Ministry of Corporate Affairs. (2026). National Corporate Social Responsibility Data Registry (FY 2023-24 Filings). Government of India National Data Portal. https://www.mcacdm.nic.in/csr-data.php
Ministry of Corporate Affairs. (2026). Annual filings by companies on development CSR expenditure totals over 1,44,159 crores in last five FYs (2019-20 to 2023-24). Press Information Bureau (PIB), Government of India. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2226018&reg=3
Ostrom, E. (2009). Beyond Markets and States: Polycentric Governance of Complex Economic Systems (Nobel Prize Lecture Essay). Royal Swedish Academy of Sciences. Open-access text available via the official Nobel Prize Archive Repository: https://www.nobelprize.org/uploads/2018/06/ostrom_lecture.pdf
Sattva Consulting. (2025). India CSR Outlook Insights. Sattva Ecosystem Research Briefs. https://www.sattva.co.in/csr-outlook-2025/
The Institute of Company Secretaries of India (ICSI). (2021). Company Law Corner Statutory Bulletin (Companies CSR Amendment Provisions). ICSI Enforcement Registry. https://www.icsi.edu/media/webmodules/FEBRUARY_COMPANY_LAW_CORNER.pdf

A Note on My Background

SK Saiful is based in Kolkata, India. He currently works at Fortify Health, a GiveWell-supported nonprofit incubated by Ambitious Impact (formerly Charity Entrepreneurship), focused on reducing iron-deficiency anaemia in India through wheat flour fortification as a cost-effective public nutrition intervention. Before that, he was a Research Associate at International Management Institute Kolkata. He was also part of the Don Lavoie Fellowship for AY 2023-24 at the Mercatus Center, grounded in Hayekian institutional analysis and Ostromian governance frameworks.
His research interests sit at the intersection of EA cause areas in global health and development, impactful philanthropy, and thinking through how to drive poverty alleviation through high-impact, evidence-driven interventions at scale.

Frontpage posts - Effective Altruism forum viewer

EA sexual harassment: Where’s the evidence we’re worse than baseline? by Nathan Young

On the desperate search for high-stakes action-guidance in TAIxAnimals by Jo_🔸

Context

1. Against “ro­bust­ness”

Streetlight­ing as an al­ter­na­tive?

How about mak­ing cer­tain tra­jec­to­ries more likely?

One thought on Mo­ral Cir­cle Expansion

2. En­dors­ing virtues and moral trade

3. Mak­ing small changes in per­sonal strategy

Tips on Becoming An AI Safety Generalist by Brittany Miller

Announcing the IAP Entrepreneurship Track: A New Path for Aspiring High-Impact Founders by High Impact Professionals

Why We’re Launch­ing The En­trepreneur­ship Track

Pro­gram Overview

How Does This Differ From Other En­trepreneurial Pro­grams?

Who Should Ap­ply to the En­trepreneur­ship Track

Im­por­tant Dates

Know Some­one Who Should Ap­ply?

Ques­tions?

Landmark new METR report: Can AIs already start ‘rogue deployments’ inside AI companies? by 80000_Hours

Transcript

What could an un­re­leased AI get away with? – the new METR re­port [00:00:00]

Mo­tive: Why grab more com­pute? [00:01:54]

Op­por­tu­nity: YOLO mode and jailbreaks [00:05:46]

Means: Brilli­ant idiots in data cen­tres [00:11:02]

We have to test un­re­leased mod­els [00:15:45]

Espe­cially if AI R&D is com­ing in 2028 [00:18:30]

Learn more

Where should the next EA Impact Hub be? (Take the poll by 27 May) by Moïri Gamboni

TL;DR

What the Im­pact Hub is

What Da Nang has been like

The next hub

A longer-term track

What we need from you

Who this is for

Who we are

Stay in the loop

Hiring: Head of Community Engagement US—Giving What We Can by Giving What We Can🔸

Hiring: Head of Community Engagement UK—Giving What We Can by Giving What We Can🔸

Nature’s value mostly doesn’t come from present-day benefits to humanity by Tandena Wagner

What Has Na­ture Done For Me Lately?

Break­down of Ecosys­tem Services

The Value Es­ti­mates In This Table Are Flawed

Some Other Ideas for Mea­sur­ing Value

Op­tion Value

A Case for Bio­di­ver­sity as a Biotech Re­source Library

Con­clu­sion

Ap­pendix:

Sources for my Claims about Ecosys­tem Services

Nutri­ent Cycling

Water Sup­ply

Pest Control

Wild-Caught Fish

Pollination

Disease Control

Flood Control

Water Purification

Ma­jor Assess­ments of Cur­rent Value from Na­ture:

Re­cent Assess­ments:

Older Assess­ments:

Other Assess­ments:

Announcing the Rethink Priorities Cross-Cause Fund by Rethink Priorities

TL;DR

Cross-cause pri­ori­ti­za­tion in effec­tive giv­ing is underdeveloped

Giv­ing is full of moral and em­piri­cal uncertainty

How EA has his­tor­i­cally han­dled this

Why think­ing in cause ar­eas isn’t enough

Why not just pick the best cause and fund it?

At the end of the day, in­ter­ven­tions are what’s funded — not causes

What would ad­dress the pri­ori­ti­za­tion problem

We need ex­plicit modeling

The bar­ri­ers to cross-cause pri­ori­ti­za­tion are lower than they once were

Our solu­tion: an ex­plicit and trans­par­ent cross-cause pri­ori­ti­za­tion model

The model

Giv­ing op­por­tu­ni­ties cur­rently in­cluded in the model

The Cross-Cause Fund

Who is it for?

How does it work?

Our cur­rent recom­mended al­lo­ca­tion across cause areas

1. Against “robustness”

Streetlighting as an alternative?

How about making certain trajectories more likely?

One thought on Moral Circle Expansion

2. Endorsing virtues and moral trade

3. Making small changes in personal strategy

Why We’re Launching The Entrepreneurship Track

Program Overview

How Does This Differ From Other Entrepreneurial Programs?

Who Should Apply to the Entrepreneurship Track

Important Dates

Know Someone Who Should Apply?

Questions?

What could an unreleased AI get away with? – the new METR report [00:00:00]

Motive: Why grab more compute? [00:01:54]

Opportunity: YOLO mode and jailbreaks [00:05:46]

Means: Brilliant idiots in data centres [00:11:02]

We have to test unreleased models [00:15:45]

Especially if AI R&D is coming in 2028 [00:18:30]

What the Impact Hub is

What Has Nature Done For Me Lately?

Breakdown of Ecosystem Services

The Value Estimates In This Table Are Flawed

Some Other Ideas for Measuring Value

Option Value

A Case for Biodiversity as a Biotech Resource Library

Conclusion

Appendix:

Sources for my Claims about Ecosystem Services

Nutrient Cycling

Water Supply

Major Assessments of Current Value from Nature:

Recent Assessments:

Older Assessments:

Other Assessments:

Cross-cause prioritization in effective giving is underdeveloped

Giving is full of moral and empirical uncertainty

How EA has historically handled this

Why thinking in cause areas isn’t enough

At the end of the day, interventions are what’s funded — not causes

What would address the prioritization problem

We need explicit modeling

The barriers to cross-cause prioritization are lower than they once were

Our solution: an explicit and transparent cross-cause prioritization model

Giving opportunities currently included in the model

Our current recommended allocation across cause areas

Methodology behind our model and tools

Fund-by-fund cost-effectiveness estimation

Defining worldviews

Outputting an allocation

Future cross-cause prioritization plans

Support the research behind the cross-cause prioritization model

Conflicts of interest statement

Animals are not one issue

The problems are staring right at us

Stop looking past them

Additional considerations

Funding amount and grant structure

Ideas we’d be excited about:

Selection criteria and evaluation process

Application logistics

A platform, not an intervention

What the platform delivers

How the platform is designed

Rethink’s way forward

How countries can get there

A closing thought

Differences in the world that motivate the change

How much of an update should we make?

Lessons.

Appendix.

B: The Startdown.

C: The Salvage.

India’s ₹35,000 Crore Corporate Social Responsibility (CSR) Mandate: An Underleveraged Frontier for Effective Giving?

The Basic Observation

What’s Actually Going Wrong with Indian CSR?

1. It’s designed as a compliance mechanism, not an impact mechanism.

2. CSR heads face the wrong incentives.

3. The NGO sector has a capacity problem that limits evaluability.

4. The intermediary layer is the actual bottleneck—and it’s largely invisible to EA analysis.