Branding AI Safety Groups: A Field Guide

By Agustín Covarrubias 🔸 @ 2024-05-13T17:17 (+44)

This article is the first in a series I plan to publish on different aspects of AI Safety group strategy. The aim is that, eventually, these articles will form the basis for a new resource center for AI Safety Groups. Note that these articles aren’t being published in any particular order.

TL;DR: AI safety groups should carefully consider their branding strategy to attract target audiences and avoid silently losing talented individuals by giving the wrong impressions. Key things to do are identifying audience profiles, tailoring the value proposition and differentiating your group from others. Common branding strategies (technical, institutional, impact, or broad appeal) each have tradeoffs to weigh based on your goals and your context. Regularly seek feedback to refine branding rather than relying solely on intuition.

Introduction

In some ways, it’s still the early days of AI Safety groups. Groups, whether they’re at universities, entire cities, or online, are still trying to figure out what works and what doesn’t. In many ways, AI Safety groups are much more heterogeneous than, say, EA groups, and this is to be expected: there isn’t broad agreement about things like group strategy, programming, or even a defined set of common goals.

One aspect of this heterogeneity is branding. This is partly a good thing: AI Safety is a field, not a cohesive community, and it often needs to appeal to very different audiences. At the same time, we could benefit from sharing lessons learned from aspects like branding, especially as groups continue to experiment and collect feedback from the real world.

This post is a provisional field guide for group organizers to think through their branding and incorporate some of the early lessons learned from other groups.

Why branding matters

First impressions matter. It’s easy for organizers to underestimate how many talented people they lose to self-exclusion. The “vibes” you project can be a dealbreaker for many people, and these are often silent failures: it’s difficult to realize this is happening unless you make an explicit effort to find out about it.

Many group organizers have seen steep and surprising benefits after branding changes[1], and based on my interaction with many organizers, I suspect many groups’ early struggles with outreach stem from the problems I describe below. Branding is not a silver bullet, but it certainly helps.

But before we jump in, a brief disclaimer: branding is not about being strategically dishonest or hiding your intentions from others, it’s about communicating clearly and making a compelling case for why others should care about your group. It’s an exercise in communicating clearly, not in deception, and it’s important to distinguish between these two things.

What failure looks like

To illustrate my point, here are some examples of failures that can occur from using the wrong branding strategy for a particular audience. All of these are loosely inspired by actual experiences from groups:

You get the gist: it’s important to consider how you’re perceived by others. Particularly those in your target audiences, but occasionally those outside of it.

Edit: Note that these can also be successes, depending on your target audiences. Signaling clearly who you're not looking for is also valuable, as it avoids bad experiences from misaligned expectations about your group.[3]

Defining your strategy

These are some things to consider when rolling out your own branding strategy. Even when a group is following or copying an existing strategy (as the ones below), I think organizers would benefit from taking some time to think through these considerations anyway:

If your group changes scope (for example, by starting an AI Governance fellowship in an otherwise technical group), you should probably revisit these considerations to ensure you’re not failing to update your strategy for a different set of audiences.

It’s easy to dismiss as unnecessary distractions from organizing, but once you let organizing flow downstream to some of these considerations, you’ll find that many things become much easier.

Some high-level strategies

I’ll review some high-level strategies groups might want to follow and their tradeoffs, giving examples of each. This is not meant to be an exhaustive list, but I suspect it encapsulates the implicit strategies of many groups relatively well.[4] 

Note that these strategies are ultimately context-dependent. Different cultures, countries, or even universities might interpret certain elements differently, and you should be ready to adjust accordingly. That said, in my experience, new organizers frequently over-update to their local context[5], so a good strategy is to copy other strategies at first and then tweak them based on actual feedback, not just intuition.

You don’t necessarily need to choose only one of these. Groups might benefit, for example, from using a broad-appeal strategy for the group itself while branding their technical and governance programs differently. If you have a website, your landing page can use a strategy different from the page for your technical fellowship, as long as the group’s brand still feels relatively cohesive.

Technical-leaning

The website of the AIS Initiative at Georgia Tech reflects a technical-leaning strategy.

Institutional-leaning

The website of the Yale AI Policy reflects an institutional-leaning strategy.

Impact-leaning

The website of Stanford AI Alignment reflects an impact-leaning strategy.

Broad disciplinary appeal

The website of Harvard's AISST reflects a broad appeal strategy.

Conclusion

AI Safety groups have great potential, but this potential can sometimes be unnecessarily curbed by practical considerations, such as how they decide to brand themselves. Hopefully, this guide can serve as a starting point for organizers, both new and experienced, to think through some of these considerations and iteratively improve their groups.

Annex: Some practical tips on websites

If you have a website, try to make sure that:

If you want a website, but you’re not confident in being able to afford/build a custom one:

  1. ^

     The most salient example of this is the big jump in applications to groups that moved from an EA-focused branding to an AIS-focused branding. See here for details.

  2. ^

     The issue doesn’t seem to be that EA is involved, but rather that none of the elements that would appeal to a technical audience alone are present. This doesn’t mean you should hide your group’s involvement, but rather that you should think about what your audience cares about in the first place.

  3. ^

    Thanks to @Jamie Bernardi for suggesting this addition

  4. ^

     Some groups which seem hard to encapsulate in one of these: AI Safety Tokyo, UVic AI and WAISI. I expect some of the examples to be a bit of stretch, since there aren't good prototypes for some of these strategies .

  5. ^

     One example is that when I started my first EA group, the other organizers and I assumed that it would be better to figure out most things by first principles, given the different cultural context at a Chilean university. It took us a year to realize that many of the recipes followed by other groups actually transferred pretty well, and ours lacked real experience.

  6. ^

     This might seem like a contradiction with appealing to AI ethics, but I think branding is separate from content, and groups pursuing this kind of framing can appeal to both, either or neither.

  7. ^

     This isn’t to say this is necessarily a bad thing, but exclusivity can also be a powerful signaling tool. Deciding where to lay in the spectrum between openness and exclusivity is one of the trickiest decisions when it comes to group strategy.


Ulrik Horn @ 2024-05-16T03:23 (+4)

I really like that you have both highlighted the importance of branding as well as made some reflections on how it affects different groups. I also really like the suggestion to think about one's audience. On the part about audience I think one should also think carefully about what one actually needs - jumping straight to interests of your audience without considering what you actually need could introduce bias and ineffectiveness into your outreach. For example, perhaps what you need is someone really good at research management, or interfacing between technical and policy people (probably not the best examples). If you go straight into interests, you might think "oh it's someone hackery that plays D&D on Friday night and should be able to make a 10am meeting on Saturday" (again, probably bad example). But in reality, maybe the people doing best in such positions are more likely to socialize on Fridays and not wake up before 3pm. 

On the same topic I also feel a bit unsure about the "grouping" above of People with a strong interest in CS or ML and People who relate to hacker culture. I loved my CS courses but never related much to hacker culture and know many like me. That said I do know there was some CS hiring algorithm that used engagement with certain Manga websites as a strong predictor of coding skills but I would be quite careful in using such correlations, and again really nailing down the "complete" set of characteristics you need, which might not only include coding skills but also communication, management, etc. I think others have written about and fleshed out in quite a lot of detail the specific needs of AI Safety and would suggest to really engage with the stated needs of the field to make sure your branding aligns with what is predicted to be specific talent gaps going forward.

AgustĂ­n Covarrubias @ 2024-05-17T21:26 (+4)

I agree that it's important to consider both needs and interests. Ultimately, a branding strategy should be embedded in a larger theory of change and strategy for your group, and that should determine which audiences you reach out to.
 

Regarding the latter, I agree that an interest in, say, hacker culture, does not adequately describe all people interested in CS. It might actually leave out a bunch of people that you should care about joining our group. At the same time, branding is all about tradeoffs, and you have to pick which things you cater to. Spread too thin, and you risk making the content too unappealing.

Ulrik Horn @ 2024-05-18T02:40 (+2)

Yeah or maybe you could do like with toothpaste: one for white teeth, another for good breath. I think it's the same toothpaste in both tubes.

tzukitchan @ 2024-05-13T17:25 (+2)

glad someone wrote up a post on this! these look like your best guesses, i was wondering if we possibly have any empirical data that can support these claims? 

AgustĂ­n Covarrubias @ 2024-05-13T17:31 (+3)

It's hard to give empirical data on this because I don't think we have a good track record of actually collecting it. I would be curious about groups trying things like A/B tests to refine their strategies.

So yeah, most of this is backed by a mix of anecdotes from organizers, marketing know-how, and some of what I learned running my old AI Safety Initiative. This is why I want to emphasize that most of this is to be taken as provisional rather than the final word on the matter.

SummaryBot @ 2024-05-14T13:01 (+1)

Executive summary: AI safety groups should carefully consider their branding strategy to attract target audiences and avoid losing talented individuals by giving the wrong impressions.

Key points:

  1. Branding matters for AI safety groups to communicate clearly and make a compelling case for why others should care about the group.
  2. Groups should identify target audience profiles, tailor their value proposition, and differentiate themselves from other groups.
  3. Common branding strategies include technical-leaning, institutional-leaning, impact-leaning, and broad disciplinary appeal, each with their own tradeoffs.
  4. Groups should regularly seek feedback to refine branding rather than relying solely on intuition.
  5. Practical tips for websites include having a clear description of the group, what it does, and how people can participate on the landing page.

 

 

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.