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

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:

1. Concentrate on the few plausible scenarios where there is intervention potential and forget the rest
2. Take individual virtues seriously and seek out moral trade with other stakeholders - maybe even by drafting your own personal code of conduct
3. 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:

1. 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.
2. 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.
3. 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!

1. ^{^}

   One of my favorite EA quotes, would make a great t shirt

2. ^{^}

   If you disagree, leave a comment down below + smash that subscribe button

3. ^{^}

   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.

Show all footnotes