AGI and Animals

This post is inspired by an Effective Altruism Forum debate week on the statement: “If AGI goes well for humans, it’ll probably go well for animals.” I’ll argue for its truth. While the week has long passed, I’ve abstained from reading its posts to first get down my unswayed thoughts.

Introduction

While impact on animal welfare carries much variance on differing AI trajectories, and I personally expect transformative AI in short order, I assume here that transformative / powerful / general artificial intelligence has the effect of speeding up progress such that advancements we would have expected in the coming decades come much sooner (basically the Machines of Loving Grace definition); how much sooner, and how quickly this transformation occurs, is out of scope. This will be, then, “AGI [going] well for humans.” One scenario that fits under this umbrella but is contrary to what I lay out here is a fast-takeoff space-colonizing singularity, which may be very bad for animals; see Bentham’s Bulldog on this.

Given a “speed-up” transformation, we can look to what we might’ve expected for animal welfare in non-AI timelines for the coming decades and extrapolate from there. This simple analysis leans towards the positive outcome.

Do human values consider animal welfare?

Current efforts toward aligned AI are concerned with either the technical alignment problem (instruction-following) or the value alignment problem (mimicry of human values, e.g., following constitutions that center human interests).

A technically aligned AI precisely following our wishes, even with safety mechanisms for human harm, would assist in procuring our food. A fully AI-managed economy would not only continue the practice of factory farming, but do so more efficiently, with whatever harms that may entail. This category of harm seems the most prominent raised by AGI x Welfare skeptics, and should be taken seriously. However, I don’t expect it to push the utility of AGI for animals into the negatives, given the acceleration examined in the next section.

Under the value model of alignment, we can look to human values to determine how powerful systems might value animal interests. Clearly, we don’t treat animals well. Tse et al., in their paper arguing for including animals in alignment, collect some statistics which lay this out well enough, so I won’t go searching:

The extent of our disregard for non-human animals’ welfare in our human values is shown by how we choose to treat them currently and throughout human history. For instance, through factory farming, we currently kill and cause immense suffering to more than 70 billion land animals (FAO, 2023; Orzechowski, 2024) and 130 billion farmed fish (FishCount, 2025) each year. All this is completely unnecessary, as it substantially reduces the quantity of food available for human nutrition.

These facts are the revealed preferences of mankind on animal welfare, and AI systems trained to act like us will share them without intervention. Vast sums of training data tell the story that killing for food is acceptable. 2020 was Peak Vegan: late-internet, and a blip in the history of written work. And still, adoption hovers only around two percent. While great moral failures of the past have been settled, banished to the “mistakes of history” pile, and endlessly disavowed, we remain far away on the question of animal rights. Thus, AI aligned to current human values (those revealed through our actions and prevailing morality) will necessarily not be aligned to animals.

It’s possible that a being aligned purely to the purported values of humanity would, in fact, come to the conclusion that harming nonhuman animals is wrong. Core values of equality or harm avoidance could lead naturally to consideration of harm to nonhuman animals, but fail due to the lack of perceived moral status. How this could go well, I’m not sure. Something close, perhaps, is Yudkowsky’s alignment proposal, Coherent Extrapolated Volition: “Rather than attempt to explicitly program in any specific normative theory (a project which would face numerous philosophical and immediate ethical difficulties), we should implement a system to discover what goals we would, upon reflection, want such agents to have.” He expects that

a vegan or animal-rights activist should not need to expect that they must seize control of a CEV algorithm in order for the result of CEV to protect animals. It doesn’t seem like most of humanity would be deriving huge amounts of utility from hurting animals in a post-superintelligence scenario, so even a small part of the population that strongly opposes* this scenario should be decisive in preventing it.

The closest inkling we’ve gotten in the current iteration of things, in Claude’s Constitution: Anthropic is concerned with “physical, psychological, financial, societal, or other harms to users, operators, third parties, non-human beings, society, or the world” [emphasis added]. Later: “Claude has to weigh . . . [the] welfare of animals and of all sentient beings” against other values.

---

If AI systems end up being more intelligent and capable than humans, we hope that they will value us, and align them with this goal in mind. We can generalize towards the principle that it is good for more intelligent beings to value less intelligent beings. Perhaps we’re acting only in self-preservation and this doesn’t reflect a moral principle, but it should prompt reflection. Human supremacist attitudes ultimately align with the anti-humanist accelerationist view of AI as a successor species if, in fact, might makes right: Humans were the inheritors of the earth, lucky to be the highest intelligence in existence; we may lose that status in the near future, and that that is The Way of the Universe. Do we deserve to be treated as equals? Alternatively, we can imagine a world in which all sentient beings are valued, including nonhuman animals and even artificial minds.

Expected harm reduction

Even without alignment to animal interests, technological advances can improve welfare. While the jump from traditional to factory farming and its successive improvement has been a disaster in that regard, I think we’re in late-stage suffering. In-ovo sexing reduces chick culling; controlled atmosphere stunning reduces suffering during slaughter.

But one technology stands clear above the rest in its potential: cultivated meat (obviously).

Prices have come down dramatically already, but cost parity is possible. But the cultivated meat industry is having financing issues: Believer Meats ceased operations in December 2025 after completing its $120 million North Carolina facility. Meatable and SCiFi Foods shut down. This follows a broader alternative protein downturn. Beyond Meat (now Beyond) is attempting a pivot, since it’s “just not the moment for plant-based meat right now.” If input costs can be brought down with advancements in automation, we could get over the investment hump and deploy at scale.

The gains here would be so massive that they outweigh the potential increase in factory farm efficiency. Bryant et al. (2019) found the portion of respondents very likely to regularly purchase “clean meat” in the US, India, and China to be 29.8%, 56.3%, and 59.3%, respectively. For the most opposed, price is not the limiting factor. But for those on the fence, and to turn the ‘very likely’ into actual customers, price remains important. If even a fraction of that group substitutes factory farmed meat with cultured, the reduction in suffering would match an unheard-of wave of vegetarianism.

As with any potentially rapid advance in areas entangled with regulatory oversight (the canonical example here is drug development), governmental intervention may prove to be the largest blocker. In May 2024, Florida became the first state to ban cultivated meat, a move recently upheld by the 11th U.S. Circuit Court of Appeals. Five more states followed, with others implementing moratoriums or barring the use of public funds for the research or purchase of cultivated meat. The main objections seem to come from the agricultural industry, though they’re supplanted with health concerns.

The reactionary backlash against even the beginnings of progress in the area points to an opt-out problem. There will always be a contingent of people who don’t trust food “made in a lab” - it’s unnatural (alas, that’s what makes it a feat of human progress). We can only hope (and advocate) to convert as many as we can; thankfully, this ask seems like it could be much easier than the prior one to give up meat or animal products entirely.

A culture speed-up?

While we can expect culture to lag behind the rate of technological progress, simultaneous breakthroughs in our understanding of the world could usher in a similar speed-up effect on culture. One interesting path towards increased public concern for animal welfare is a stronger understanding of consciousness. While this could come from new superintelligence-conducted research on these questions in particular, something like the hard problem may be fundamentally out of our grasp.

Interpretability work could also play a role. If large progress is made here (likely, given the scenario, since AGI has gone well for humans), we’ll have an increased understanding of the inner-workings of models’ abilities. Attempts could then be made to apply interpretability techniques to biologically-conscious beings. How feasible this is, I’m not sure. Dario, in Machines of Loving Grace, points to a study of the mouse visual cortex rediscovering vision model mechanisms as an example of the potential of interpretability research for neuroscience.

There is some signal that fuller awareness can translate to welfare action: The UK Animal Welfare (Sentience) Act 2022 extended legal protections to cephalopods and decapod crustaceans after a government-commissioned review of evidence for octopus and crab sentience. But the entrenched interests present in food and animal agriculture present a much stronger opposing force than in narrow cases. And even with a limited understanding of consciousness, there has been broad consensus that neurological substrates of consciousness are widespread among nonhuman animals, that they have the capacity to suffer. If this hasn’t made for a step-change in public thinking around welfare, what could? It would have to be revolutionary.

The role of animals in a transformed future

If transformative AI is far away, we should hope that human values evolve to reject speciesism. Then, human-aligned AI would value animal welfare by proxy. Whether value progression will matter in a short-term transformation scenario depends on how AI will adapt to changing human preferences. If there is a lock-in effect in which we “the values determining the future of earth-originating intelligent life” are set and irreversible at the time of an explosion in machine intelligence, having the right values at that moment similarly explodes in importance. Any consideration of the welfare topic with this effect in mind should lead us toward the conclusion that lock-in is probably really bad. We want to preserve the ability to make moral progress - thus, we should leave out specific attitudes around meat consumption, for example, and hope that, as humans may come to reject the practice, then too will aligned AI.

Reduction in suffering is the baseline, and the goal I’ve focused on here. But we might place value also in the continued existence of unique species, in having a broad array of life on earth, even if a world without nonhuman animals is solved for. An additional space of trade-offs presents itself; if we end up covering Earth in compute, destroying habitats, animals could be relegated entirely to zoos. But such an event would solve also for wild animal suffering, an enormous wellspring of harm I’ve not addressed here. All these points I simply touch on; they point toward further complexity at the intersection of transformative AI futures and animal welfare that are worthy of future analysis.