Context

Skip ahead if you’ve read my other post on transformative AI and animals. If you haven't, this section may answer some additional questions.

Confidence level: this was timecapped and may be rough in places. I put low confidence in my judgment on speculative issues. This is an offshoot of a project I did during the Summer 2025 FutureKind AI Fellowship.

Intent: It seems that transformative AI x Animals (henceforth TAIA) is bottlenecked by a lack of object-level interventions.^[1] This is a list of interventions in transformative AI x Animals that could make for feasible pilots (by feasible, I mean something that could be done by 2 FTEs with a year of funding, like an Charity Entrepreneurship-incubated charity). This excludes most meta-work like running fellowships and organizing conferences.^[2] Most interventions here have been suggested in pre-existing posts: this is crowdsources rather than brainstormed. 

Scope: These interventions don't have to appear positive to be on the list. Many of them are already being implemented, and they’re not all animal-focused. Focusing on TAIA means choosing interventions that are grounded in scenarios where AI is transformative and probably changes our intervention levers. I don’t cover interventions that already seem promising without transformative AI, like improving welfare in precision livestock farming. I also don’t cover artificial sentience, which I consider to be a separate issue.

Hedging on timelines: While I made this list with timelines to TAI of 7 years or less, I don't think that this changes much. You may be more excited about meta-work, but even then, staying meta for too long could be negative for the cause area. This post has broad agnosticism on what sort of TAI trajectory is most plausible. Your judgment on this will influence what interventions you think are good.

Other caveats:

• I overall assume that good interventions are hard to find (DiGiovanni 2025).^[3] I suggest that readers don't over-update from me not finding any specific intervention positive.
• I notice that many of these are only tractable if AI is developed in the West.^[4]
• Format: the bullet under the presentation of the intervention is my superficial judgment.

Interventions

The bullet under each description reflects my first-pass judgment of the intervention.

Judgment Calls on How Present Issues Can Influence TAI

My judgment: these interventions look somewhat reactive and the case for them doesn't depend on rigorous prioritization. However, they benefit from clearer feedback loops - though it's not certain that they'd pay off in AI-transformed futures.

Developing AI-Enabled Epistemic and Coordination Tools

This involves building or deploying AI tools designed to improve human decision-making, deliberation, and compromise. The theory of change is that better epistemic environments will allow humanity to surface and act on latent preferences for animal welfare.

• Such tools can be developed, but the causal chain to animal benefits is long and conjunctive. It has to be noted that this is conditional on some slow-ish takeoff. We don’t have a clear idea of how these tools would influence real-world outcomes. Moreover, animals are at the very bottom of political priorities: thus it’s not clear that any coordination tool that strongly considers animal welfare could be considered particularly useful, especially in periods with large-scale changes. But I don’t think this has been looked into, and maybe there are promising ideas.

Preventing, or Improving, Terraforming and Similar Processes

Advocating for space governance norms that keep space pristine from interactions with earth-originating non-human biological life.

• Ambitious and unlikely to gain traction, though a few advocates can probably find themselves in good positions, as experts or advisories, to have a very marginal influence on discourse, and, with more speculation, future decisions. Obviously, this won’t influence rogue TAI much. The major problem is that interventions there can easily backfire by raising salience, or throwing this issue into public discourse too early. In terms of outcomes, it makes sense to keep in mind that it’s unlikely that space will remain pristine if we don’t terraform it: it’d have to be compared to alien and AI counterfactuals.

Seizing Opportunities in Current Priority Areas

Whether it is in farmed animal welfare, alternative proteins, or wild animal welfare, there has been much discussion on how TAI could offer great opportunities for leverage, if seized appropriately. Interventions there could either mean: deprioritizing current programs and pivoting to new ToCs that bet on specific opportunities coming from AI advances (eg R&D having much faster feedback loops in alt proteins); or having one researcher per area exploring a large range of potential interventions.

• The upsides could be worth it. It somewhat hinges on the assumption that post-TAI futures won’t be too weird and that humans will still have power, so it’s not clear that seizing opportunities early would make a big difference. Feedback loops would still be a problem for the most part, making tractability lower than one could initially expect. Given that this would be similar to preexisting interventions, second-order effects would probably be similar to what we already know about.

Judgment Calls on How TAI Trajectories Affect Animals

My judgment: We're far from reaching a consensus on what AI safety interventions most effectively achieve their goals, or even what the priority goals are (especially from the perspective of impartial welfare). Thus, crucial considerations will be subjective. 

Improving the Perception of Animal Welfare in Influential Spaces

Building credibility within tech circles through high-signal engagements, and giving a professional, transpartisan image of animal welfare. Lewis Bollard’s excellent appearance on the Dwarkesh podcast was a highlight, but there are other, smaller-scale examples of this.

• Probably dominated by a few good opportunities, but at least this wouldn’t only be supported by people who buy that TAI is important (though maybe that should make us suspicious?). This mostly rests on the assumption that TAI will be influenced by human values, and developed in the West, but this still covers a variety of trajectories. The most fragile links are: TAI retaining such values, or such values not wrongly prioritizing certain animals.

AI Pause / Anti-AI advocacy

If no outcome with transformative AI (save for extinction) avoids the worst harms to animals, a logical focus could be pausing AI development, banning superintelligence, or developing strict red lines.

• It’s obviously feasible: advocacy groups like Pause AI are looking for volunteers, and some AI governance orgs have work that falls into this category. Of course, this will depend on your judgment of how feasible it is to restrain the development of TAI - there’s no shortage of content debating this topic. The track record of public-facing advocacy, however, may be a neglected consideration.

Support Broad AI Safety Efforts

This approach involves contributing to mainstream AI safety work to ensure TAI “goes well”. A lot of animal advocates have done this already.

• Feasible and scalable... Nonetheless, all other aspects of the assessment rely on whether one thinks that desirable outcomes in the eyes of the AI safety consensus are better for non-human animals; and perhaps even less consensually, on whether you think the AI safety field is currently guiding the world closer to these desirable outcomes.

Preventing Value Lock-In and Takeovers

Some assume that the status quo is so bad that a lock-in of current (or worse) values is the most important thing to prevent.

• There are orgs working on this (Forethought, Macroscopic) that may be looking for talent or funding. However, judgments on lock-in often feel rash, as it’s very hard to have a counterfactual idea of what a constantly evolving future would look like. Then there’s an added issue of interventions around this being considered extremely risky due to attention hazards and possible lack of tractability.

Judgment Call on AI values

My judgment: We don't yet know much of anything about AI values, especially in different worlds. Even holding certain broad positive attitudes toward “animals” may not be robust enough to bring about the best consequences: even TAI may not be able to cover all animals’ interests, and partial consideration could be awful for the majority of animals. However, at least the ToCs of the interventions are somewhat agnostic on whether TAI disempowers humans (while retaining certain values which humans built into it - which seems strange, especially in the long run); or whether humans remain in control.

Integrating Animal Welfare into AI Governance

Lobbying for the (minimal) inclusion of sentient nonhuman interests within intergovernmental frameworks, codes of practice, etc. This has already been done to an extent in the context of the EU AI Code of Practice. This could probably be done by pre-existing orgs.

• There’s at least one example of such a thing being done, which is good. However, there aren’t that many commitments of the sort yet, and it may continue to be the case. Even if there turns out to be more opportunities like that, starting a new org for a lobbying purpose may somewhat backfire by seeming like the work of a narrow interest group, or encouraging more and more asks in AI regulation, which could dilute lobbying efforts. However, this effect may be smaller in expectation than the win.

Corporate Commitments & Benchmarks

Encouraging AI labs to include "sentient beings" in training specifications, mission statements, and "constitutions" to explicitly protect animal interests. Organizations like Sentient Futures are already exploring this, and some researchers appear receptive to these framing shifts. Progress could be made through scaling these orgs (or through ambitious moves by those already in the field), or through founding a new, more focused, 1 or 2-person org.

• While some labs are open to the idea, securing a concrete success remains difficult and dependent on timing. There is a risk that this encourages more ethical "asks" from other group that makes labs less collaborative, but this marginal effect doesn’t obviously outweigh the potential for a success.

Targeted Training Data

This intervention is being pursued by Compassion in Machine Learning (CaML). They generate synthetic pretraining data oriented toward consideration for animals (and sometimes digital minds).

• Already underway. More importantly, CaML reports experimental evidence that this can improve animal-compassion scores in open-weight models, with effects that persist after standard fine-tuning. It remains to be seen whether this synthetic data would be used by labs, but it seems passably likely.

Animal-Inclusive Alignment Research

Starting a technical (or even foundational?) research organization to identify potential animal-harming tendencies in AIs. Something that is close to this is what CaML is doing, but it seems that many interventions and ToCs in that style could be worth pursuing. Foundational research may be harder with fear of short timelines, but there may be interest in thorough research on whether alignment to “all sentient beings” can ever be robust to the first objections that have been raised against it.

• One of the hardest to launch, as it may need competitive salaries and a hub position. Moreover, there’s no very clear progress on the main large-scale alignment issues, in spite of hundreds of researchers working on the problem.

Interventions That I Find Harder to Defend

My judgment: I currently don't see why one would pursue these if they want to improve outcomes for animals.

Crazy Train Moral Circle Expansion

Some advocates may want to seize unprecedented societal shifts to raise the salience of animal sentience. If AI actually enables communication with some non-human animals, this will almost inevitably be used for narrative-building. Some have suggested using empathy towards human-like LLMs for the same purpose.

• I think this will almost inevitably happen, so it’s already unlikely to be the best intervention to support at the margin. I also think that it raises many issues, such as the difficulty of gaining traction with new narratives when times will be changing fast. More importantly, simply getting people to acknowledge that “all animals matter” doesn’t seem like such a strong strategy to help animals. Indeed, even within the small circles that prioritize animals, it’s hard to get people to take smaller, weirder animals seriously. Within society, increased empathy toward larger animals might be a minor factor of the small animal replacement problem. And finally, animal communication is likely to make this worse, as we’ll probably talk with whales long before we talk to shrimp.

A Patient Philanthropy Fund for TAIA

Getting funders who think TAIA may be important to pool resources in a fund in order to act when we have better indications of trajectories, and when some TAI scenarios can be falsified (inspired by Founders Pledge).

• This has the benefit of avoiding two issues in TAIA: prioritizing interventions that are conditional on trajectories that don’t come to be, and the possibility that humanity never even develops TAI. (I assume the funds could go to conventional animal welfare interventions if this ends up “falsified”. However, feasibility seems very low given how nascent the field is. It would also probably be very hard to know when to act, as we have no track record on this situation.Mass Movement Building and Public Pressure

Sparking Public Discourse on AI-driven Harm to Animals

Tactics could include getting high-shock stories in the media (e.g., PLF abuses or autonomous vehicle accidents involving wildlife), and trying to draw the line from that to future larger-scale AI harms. The theory of change would probably be that if there was more pressure for AI to not harm animals, this would improve the development on the margin.

• Aside from the fact that gaining traction on this with 2 FTE would be hard, this would be highly likely to backfire, and the message would be low-fidelity. Given that current frontier AI models are somewhat animal-friendly in their stated values, something stands to be lost by spreading the message that AI is anti-animal. Though of course, second-order consequences of actions being what they are, this could turn out to be a positive strategy, but I wouldn’t bet on it.

Searching for a "Cause X" within TAIxAnimals

A focused research effort to identify a single, high-leverage "priority issue" that could serve as a focal point for the movement’s limited resources. Could be done by one independent researchers or members of a think tank.

• Independent research without feedback loops seems to have little values, and there’s not much of a track record for finding actionable causes “out of nowhere” in the past years. Even if a somewhat promising novel intervention came from there, it’s hard to know whether the infrastructure to implement it would exist. Given that there’s already a substantial amount of meta-work, more focused research doesn’t seem like a current priority.

Shallow takeaways and next steps

Nothing in the list emerged as strongly compelling, but I'm generally on the skeptical side. Given that launching misguided interventions could be irreversibly costly for a new cause area, it makes sense to do some red-teaming, to preserve option value and not waste resources. 

I wouldn't be surprised if some readers of the post disagree with most of my judgment. If so, I'd love to hear:

1. Why you think some interventions here are robustly positive
2. If you think I’m missing some robust interventions
3. If you have plans to implement such an intervention in the future

A few weeks ago, Aidan Kankyoku from the Sandcastles blog) did something similar, and even longer, in the context of ending factory farming.^[5] Since his takeaways are different from mine, you may appreciate reading it.

Acknowledgements

The initial idea for a list of intervention ideas came from Max Taylor, which does not mean that he endorses my suggestions or conclusions. Kevin Xia gave quick feedback on an earlier version of this list. Many of these interventions were implicitly suggested in previous posts on AI x Animals. LLMs were used to polish my own notes on most interventions. 80% of it was rewritten, some sentences were kept the way ChatGPT rewrote them.

1. ^{^}

   I don't think that resolving bottlenecks in TAIA is a priority, but I assumed this could be a good nudge to those who think differently.

2. ^{^}

   I have been inconsistent with the criteria in order to include interventions which I know are being considered.

3. ^{^}

   The implementation / outcome robustness framework justifies this well in my view:
   Outcome robustness: Intervening on X in a given direction would be net-positive. 

   Example of failure: It’s unclear whether “human” space colonization (SC) is better than misaligned AI SC, given how many systematic ways these coarse categories could differ in various directions. (Especially when we consider exotic possibilities, like interactions with alien civilizations and acausal trade.)

   Implementation robustness: Our intervention on X would (i) change X in the intended direction, and (ii) avoid changing other variables in directions that might outweigh the intended positive effect.

   Example of failure: AI safety interventions might (i) increase the risk of human disempowerment by AI, e.g., by increasing AI companies’ complacency; or (ii) increase the risk of extinction by causes other than successful AI takeover, e.g., a great power war with novel WMDs.

^{^}

This doesn't mean that I endorse it, it's just a blind spot.

^{^}

"Crazy Train Moral Circle Expansion" was inspired by a section in his post.

Show all footnotes