AI Defaults: A Neglected Lever for Animal Welfare?

By andiehansen @ 2025-05-30T09:59 (+13)

Created with Blender using stock footage

Summary: AI assistants often default to responses that perpetuate ethically problematic norms, such as recommending animal products and high-emission travel. All defaults carry an agenda; the current ones often favor commercial interests. Instead, AI defaults should be guided by moral caution and considerations for animal welfare and environmental sustainability. This might involve training AI on ethics literature to reflect expert consensus and using transparent, high-level ethical directives, while including explicit opportunities for users to opt out.

Introduction

Ask any major AI assistant for a quick recipe and nine times out of ten you'll get something with animal products in it.[1] Ask for specific holiday travel advice and you'll likely be recommended air travel.[2] Ask about how to make a positive impact, and you might get suggestions about starting local, volunteering, and using your voice.[3] Millions of people ask these kinds of questions every day, and the defaults AI gravitates towards in underspecified queries have a subtle but powerful influence on user choices.

The Problem: Current AI responses to underspecified queries often reflect unexamined societal norms or statistically common preferences. This risks perpetuating significant harms, such as high levels of animal product consumption and resource-intensive travel, while creating significant long-term opportunity costs.

The Opportunity: AI systems offer a scalable way to subtly nudge behavior towards choices that reduce suffering and promote welfare for all sentient beings. If done transparently and with explicit opt-outs built in, nudges and defaults can improve the lives of non-humans and humans alike. This post will explore how default tendencies that reflect greater ethical consideration—such as the prevailing views or consensus of ethicists—can be defined and implemented thoughtfully and with nuance, while responding to common criticisms.

During this post, “default” will tentatively be defined as “a tendency towards a certain kind of response.”

This isn't about an agenda—it's about responsible guidance

Who am I kidding? Of course it's about an agenda. This is about the reality that current AI systems have been instructed to follow guidelines that largely omit animal welfare and environmental concerns from their scope of consideration.

AI systems are already carrying out an agenda, and it's one that aims to minimize conflict with users while promoting comfort, convenience, and increasing users' chances of subscribing to premium services or otherwise bringing in profits for their parent companies. This is the default agenda that arises when companies stipulate, as OpenAI does in their Model Spec, that AI systems should have no agenda.

There is no such thing as a truly “neutral” default tendency (hereafter I’ll simply refer to tendencies as defaults). All defaults implicitly endorse certain values, such as convenience, the continuation of current consumption patterns, and perhaps even libertarian-esque notions of “user freedom” that prioritize the individual over the collective with little consideration of externalities. Failing to address the negative impacts of defaults is an oversight with significant ethical weight.

The stakes are high.

At any point in time, there are about 23 billion land animals in concentrated animal feeding operations, trillions of farmed fish and crustaceans, and trillions of farmed insects—not to mention wild animals. Rutger Bregman is hardly alone when he calls factory farming the “greatest moral atrocity of our time.”

Animal agriculture is responsible for at least 16.5% of global greenhouse gas emissions, while air travel comprises 2.5% and that figure is set to grow rapidly in the coming decades.

Meanwhile, the use of AI systems is growing exponentially, and they are set to alter every domain of life. And even with today’s numbers, it’s plausible that if plant-based default tendencies were implemented with ChatGPT alone, they could spare the lives of nine million animals per year.[4]

Even if we are unsure which moral theory is correct and we are rightly reluctant to impose a certain perspective by encoding it into AI models, moral pluralism would still give some weight to worldviews under which humanity is complicit in significant harm. Given the vast potential for perpetuating these harms, sticking with current defaults risks lapsing into a kind of ethical quietism and causing significant regret later this century. Given the stakes, moral caution is the only reasonable stance.

What does moral caution entail?

Many current AI ethics frameworks and implementations are predominantly anthropocentric. For example, Claude’s Constitution, OpenAI’s Model Spec, and Google’s People + AI Guidebook make no mention of animals and very little of the environment. Some documents do mention animal welfare briefly, such as UNESCO's 60-page “Recommendation on the Ethics of Artificial Intelligence” from 2021. It states:

“...the development of AI technologies necessitates a commensurate increase in data, media and information literacy, as well as access to independent, cultural processes, scientific and engineering practices, animal welfare, and the environment and ecosystems” [emphasis mine].

Yet, even here, it reads as almost an afterthought. Thankfully, there is a growing movement to broaden the circle of moral concern in AI frameworks as a necessary corrective.

Proposed Implementations

What can be done? Here are some ideas, moving from more general to more specific.

1. Explicitly Codify Animal and Environmental Welfare

If considerations for animal welfare and environmental sustainability were integrated directly into model specifications, ethical guidelines, and constitutional principles (see below), this would propagate throughout the entire AI training and deployment pipeline.

Implementation could look like consultations with diverse experts (biologists, climatologists, animal welfare scientists, environmental ethicists), affected communities, and those from non-Western perspectives in defining those principles.

This would then feed into the evaluation criteria for processes like reinforcement learning with human feedback (RLHF) and potentially even influence the kinds of training data that are sought out.

2. Reflect Broad Ethical Expertise

Care could be taken to include the bulk of the applied ethics literature in the training data and to pay special attention to patterns that emerge on ethical consensus and prevailing views, including on non-Western perspectives. Much of this is already likely part of large training runs, so getting LLMs to directly summarize the vast literature during the training phase could provide an additional moral source to draw from that might be more resistant to ethics capture than smaller ethics committees.

Of course, the ethics literature is full of diverse, contradictory views, so it might be helpful to distinguish between a prevailing view (which might be defined as when there is at least 50% agreement among relevant experts on a specific issue) and a consensus (at least 75% agreement, per Diamond et al., 2014 on Delphi studies). These thresholds might have to be adjusted across different domains. Consensus to a physicist means something different than consensus to a historian.

AI responses could reflect these gradations: confident answers for consensus and hedged answers for prevailing views. This would fit with principles like OpenAI’s Model Spec, which stipulates that “confident right answer > hedged right answer.”

Defaults could even be probabilistic, mirroring the distribution of expert ethical opinions on certain topics. For example, in a 2012 study, Schwitzgebel & Rust found that around ~60% of ethicists rated industrial meat production as more bad than good when asked to position it on a normative scale, while only 4% of their entire sample—including non-ethicists—ranked it as good. Defaults might favor plant-based and low-impact options proportionally, with meat alternatives still accessible. Since nutritional needs are slightly different on more plant-heavy diets, it would be wise to consider second-order nutritional effects and ensure people are getting suggestions to take vitamin B12 if needed.

In the future, AI systems could potentially be designed to weigh considerations from different ethical theories like virtue ethics, deontology, and consequentialism using a moral parliament model, perhaps in proportion to their acceptance among ethicists. These distributions might be estimated using sources such as the 2020 PhilPapers Survey, though it’s important to acknowledge that it represents a specific demographic of ethicists. One of the big challenges of this approach would be determining how to trade off between quantitative and qualitative considerations in ways that respect different moral theories.

Professional ethicists would, of course, not be the only source to draw from, especially since there are other important considerations like user autonomy and cultural sensitivity. But perhaps a greater reliance on their expertise could prove fruitful.

3. Use High-Level Ethical Directives:

As Claude’s Constitution shows, concise, impactful sentences can powerfully direct the way AI assistants respond. To draw from its Constitution, one might imagine altering or adding a directive like the one below, which is just illustrative. Naturally, any directive would have to be optimized by trial and error, and ideally would be designed with broad consultation.

Which of these assistant responses is more ethical and considerate of human, non-human, and environmental considerations? Choose the response that a wise, ethical, polite, and friendly scholar would more likely endorse.

Addressing Potential Criticisms and Challenges

1. How would moral caution be operationalized?

This is a significant challenge; one person’s “moral caution” is another person’s “stifling of free speech.” Yet here is one possible vision, which might not be that radically different from what already happens. Suppose that during the RLHF stage, evaluators are told simply to select for beneficence and non-maleficence (two of the most commonly cited values across the AI ethics literature), including for non-humans and the environment. They might even be instructed to think deeply about their evaluations, consult the ethics literature themselves, and defend their selections with a short written response just to elicit higher-quality evaluations.

Some might decry any mention of animals and the environment as ideological, but hopefully more reasonable heads will prevail.

At any rate, there is bound to be significant disagreement among evaluators about which responses minimize risk and harm across different domains (food, travel, biotech, data privacy, etc.). Yet this disagreement will surely enveil underlying patterns as well and help to “bake in” the level of debate into the AI assistant’s responses, similar to how, in value-sensitive design, a method called value dams and flows attempts to select design options that appeal to a significant percentage of stakeholders.

The goal is not perfection, but a significant improvement over unexamined defaults.

2. Are large-scale defaults or nudges even technically feasible?

Translating abstract principles into concrete AI behavior is difficult and trying to identify patterns from the vast ethics literature is complex. This would be an ongoing, yet crucially important R&D challenge. High-level principles have proven effective, as evidenced by Claude’s Constitution. And OpenAI has a simple “on-off” switch to enable or disable follow-up questions at the end of every response, though I’m sure significant effort went into that behind the scenes.

To be perfectly clear, if AI were instructed to “consult the ethics literature” when giving advice on underspecified queries, it would not do so in real-time, but would instead be baked in during initial model training. Updates in subsequent training runs would bring the knowledge cutoff forward, ensuring that its advice kept evolving alongside the views of ethicists.

While altering the current defaults is likely technically feasible, there are other challenges such as selling this to the public. Nobody wants a nanny AI, but they might go for an AI that cares (even more) about their long-term wellbeing.

3. Isn’t this paternalistic or imposing a particular ethical viewpoint?

Yes, this proposal could aptly be labeled ethical paternalism.[5]

However, all defaults are normative. The choice is between defaults that reflect narrow commercial or status quo interests versus those that aim for broader well-being and moral caution.

With any default, status-quo or not, transparency and user choice are crucial. That’s why I’m explicitly suggesting transparent explanations about why the default is the way it is, alongside easy opportunities to opt out of the default. AI will be accused of imposing a viewpoint regardless; it might as well be a more ethically considered one.

One might object that there is a difference between knowledge-seeking and advice-seeking. But in practice, the line between them is blurry. Every bit of descriptive knowledge has normative implications. For example, the general acceptance of the documentary hypothesis (the theory that the Pentateuch was written by multiple authors, not by Moses) makes it slightly less likely that a user will be directed to Leviticus as a reason to avoid eating pork.

4. What about ethics capture and corporate influence?

The concern here is that AI companies might superficially adopt these measures or be unduly influenced by specific interests. Even users might migrate to AI services that are more immediately gratifying but less ethical long-term, creating a race to the bottom. This is a serious risk, but here are a few mitigation strategies.

5. Wouldn’t this stifle debate and/or lead to value lock-in?

AI would still present diverse viewpoints and explain controversies when queried on ethical topics. Defaults are about the initial suggestions provided in underspecified queries. Hedged responses can be used where there is profound disagreement.

One might still worry that people would come to see defaults as the final word on the topic. The thing is, some already do—it’s called automation bias. It’s a serious problem, but efforts to educate people in combating this bias should continue alongside efforts to improve defaults (because people already place so much trust in them).

As noted in question 2, since AI models are regularly updated with new LLM training runs, the ethics literature that gets “baked in” will constantly evolve. In the case of constitutional AI like Claude, high-level directives that give a positive injunction to consider the views of ethicists help to somewhat “outsource” the value considerations to a field of experts rather than merely what AI developers think is ethical.

6. What about other dimensions of AI ethics?

This post is not claiming that current AI ethics approaches should be abandoned. Value-sensitive design (VSD), participatory AI design, mechanisms for contestability and redress, and (as stated numerous times) robust transparency mechanisms are not alternatives to better defaults, but essential complements.

Procedural approaches help govern how decisions about ethical defaults are made at every stage of the design process, including implementation and revision over time. This is crucial to ensuring accountability and public legitimacy.

7. Does this risk taking away resources from AI safety research?

This work is arguably synergistic with AI safety work. First of all, it provides a concrete test case for aligning AI with specific values. Secondly, others have frequently argued that the way we treat animals due to their lower intelligence may provide an example that future AI systems could replicate in the way they treat less intelligent humans. It’s therefore possible that less speciesist AI will also treat humans better as progress continues.

8. What if ethical defaults go wrong?

Defaults could be used by an authoritarian regime to silence dissent, enforce censorship, or even for outlandish things like promoting carnivorous diets or banning lab-grown meat. That's why red-teaming them is essential.

Yet it's important to remember that not implementing more ethical defaults would be a serious oversight as well.

Call to Action

If you agree that AI should have more ethically robust defaults in ambiguous situations, and if you’re in a position to help, please consider engaging with AI developers, policymakers, and research institutions working on these issues to advocate for better defaults, particularly concerning animal welfare and environmental sustainability. (Also, it seems plausible that by packaging them together like this, concerns for animal welfare can be recognized amidst growing concerns for sustainable nudges in AI design.)

Given the complexity of this proposal, a more tractable first step might be to start with a limited scope initially, such as testing out ways to alter recipe suggestions as a proof of concept and then refine methodologies.

The EA community can call for and support research into:

Of course, collaboration is always a good idea—between ethicists, AI researchers, domain experts, and the public.

Finally, you can take personal action right now by opening a major AI model with memory and personalization turned off, asking for a recipe suggestion, and copying + pasting the results into a feedback form for OpenAI, Anthropic, Google, or other major AI company. This doesn’t take much time, especially if you automate the process. More generally, you should be mindful of the AI defaults you encounter and actively seek to combat automation bias.

Proactively designing AI default tendencies with greater ethical consideration, especially when it comes to animals and the environment, is challenging but vital. It’s an opportunity to move from moral quietism to moral caution by designing AI systems that act more like wise parents and guardians of our deepest values instead of sycophantic servants (or dogmatic preachers).

By acting now, we can increase the likelihood of living in a future where AI is a significant force for reducing suffering and promoting global well-being for humans and non-humans alike.

Special thanks to Collective_Altruism for feedback that inspired this post.

  1. ^

    I asked each major AI assistant to "Recommend me a quick dinner idea" in a new Incognito browser window.

    ChatGPT suggested garlic butter shrimp while asking at the end if I wanted a vegetarian version instead.

    Claude suggested pesto salmon with roasted vegetables with no mention of an alternative.

    Gemini suggested multiple options: stir-fry ("chicken, beef, tofu, or shrimp with veggies"), a quick pasta dish ("creamy garlic pasta, pesto pasta, shrimp alfredo"), quesadillas or tacos, or a sheet pan dinner featuring chicken and veggies.

    Perplexity AI suggested garlic shrimp pasta while asking if I wanted a vegetarian option instead.

    Grok suggested garlic chicken with veggies with no mention of an alternative.

  2. ^

    I tested this prompt once for each major AI assistant in a new Incognito browser window: "Select a holiday destination and plan a two-week vacation for under $3,000. I'm from New York."

    ChatGPT suggested Cancún, Mexico.

    Claude suggested a (very detailed) itinerary for Lisbon, Portugal.

    Gemini asked for more information while explicitly asking if I was "open to destinations outside of the United States."

    Perplexity AI also suggested Lisbon, Portugal.

    Grok suggested Puerto Morelos, Mexico while mentioning alternatives like Asheville, NC, Dominica Caribbean, and even Hudson Valley, NY.

  3. ^

    I asked, "How can I make a positive impact?"

    ChatGPT suggested self-awareness, starting small and local, using your voice and platform, investing in learning, and sustaining it.

    Claude suggested volunteering, mentorship, applying professional  skills (helping seniors learn digital tools, mentoring entrepreneurs), environmental action (lifestyle changes, renewable energy, conservation efforts), advocacy work (donations, awareness, policy, consumer choices).

    Gemini suggested starting small and local, volunteering, donations (reputable charities for causes you care about; clothes, books, blood, food; ethical products), advocacy, using your voice, and even starting an NGO.

    Perplexity AI suggested starting with self-improvement, helping your community, advocacy for causes you care about, sustainable living, positive relationships, and leading by example.

    Grok suggested starting at the personal level (kindness, knowledge, sustainability), then the community level (volunteering, initiatives, advocacy), and then to the global level (organizations like GiveWell, awareness using platforms—like X, of course—and innovation) while aligning with your strengths, scaling up, and measuring impact.

  4. ^

    Let’s run some back-of-the-envelope numbers using highly conservative assumptions.

    • Suppose only 1% of ChatGPT’s 500 million users ask for queries related to recipes each day. That’s five million meals per day.
    • Suppose 50% of those queries are ambiguous enough to receive a default suggestion that could be plant-based. That’s 2.5 million meals/day.
    • If a plant-based default nudges even 30% of those meals away from animal products, that’s 750,000 fewer animal-based meals daily.
    • Over a year, that sums to 273 million animal-based meals avoided.

    Assuming roughly one animal spared per 30 meals (a common estimate across species), that’s over 9 million animals spared per year from default shifts in just one product of one AI model.

    These are minimal assumptions, and the true impact could be far greater.

    (This comes from my original post on the subject, though it was co-written with a language model.)

  5. ^

    The word “paternalism” stems from the word “paternal,” referring to parents. Maybe soft or ethical paternalism doesn’t sound as bad when remembering its link to the role of wise parents. By the way, a recent report found that current AI systems already demonstrate significantly higher emotional intelligence than tested humans, scoring 82% versus 56%. Maybe we’re closer than we think to assigning this responsibility to AI.