So, I am here, like many of you are here, because I care about animal welfare and AI welfare, and I work on animal welfare and AI welfare. I work on farmed and wild animal welfare through the NYU Center for Environmental and Animal Protection, along with our Wild Animal Welfare Program. And I work on invertebrate and AI welfare through the NYU Center for Mind, Ethics, and Policy, and in collaboration with partners like Elicit, AI research. And I work on all of these topics in part because I enjoy the variety. There are a lot of fun and interesting differences between, for instance, what cows and pigs need, and what rats and pigeons need, and what ants and bees need, and what Claude and Gemini need. And I work on these issues in part because there are similarities too.
There are parallels across them. There are interactions between them. There are some shared causes of some shared problems. And when you work across these issues in an integrative way, you can over time start to see patterns emerge. And I have found over the past two, four, six years a shared theory of change emerging for me across animal welfare and AI welfare. And this theory of change, it comes out at the margins of other talks that I give, is certainly part of some recent pieces of writing about these issues. But today, I thought I would put it front and center and offer you my current, again, broad shared theory of change across animal welfare and AI welfare. So first of all, what are some of the broad similarities that generate for me a broad shared theory of change across these quite diverse categories?
Well, the similarities are not going to be very surprising for most people in this room because we all know that these issues are important and difficult. And that right there is the key similarity that already points to a broad outline for a theory of change. Important in the conventional effective altruist sense of importance. These are issues where a large number of potentially sentient or otherwise morally significant lives are at stake. These issues are very neglected, relatively speaking. And these issues are at least potentially tractable. And these issues are connected with other very important and fast-moving issues. Animal welfare is connected with global health and environmental issues. AI welfare is connected with AI safety and alignment. And again, these issues are moving fast. And that generates a sense not only of importance but of urgency. A sense that action is needed now, or else we could accidentally lock ourselves in to a bad trajectory.
And these issues are very difficult because they force us to confront profound limitations on our knowledge, our capacity, our political will. To some extent with farmed animal welfare, but then especially with wild and invertebrate and AI welfare, we know very little about what would benefit these populations at scale. We have little to no capacity, infrastructure, institutions, resources for acting on the limited knowledge that we have. And we have little to no political will. We do not give a damn for applying the capacity that we have, and that is the fundamental predicament across these issues. They are extremely important and they are extremely difficult. And so a really important first step in a theory of change across these issues is recognizing the importance and the difficulty in equal measure. When you recognize the importance, it can be tempting to dismiss, downplay the difficulty. When you recognize the difficulty, it can be tempting to dismiss, downplay the importance. But both of those are clear failure modes. So step one is recognizing and acknowledging, internalizing, sitting with the importance and the difficulty, the extent of our responsibilities and the extent of our limitations, having the thought: we have no idea where to go and we need to start heading there now. And that is, of course, a tough thought to think. It involves some cognitive dissonance. But if you start with that thought, we have no idea where to go and we need to start heading there now, that points to a broad outline for a theory of change across these issues that I think can be helpful. And so now let me sketch the main elements of this theory of change, and then I can show you how they have informed some of our recent work on farmed and wild and invertebrate and AI welfare. So the broad theory of change that I think can be applied across these domains, despite inevitable differences that emerge in the details, is that we need to combine radical, transformational, long-term goals with moderate, incremental, short-term steps via what I think of as a dual-track and systems approach. So I can explain what I mean by each of those parts.
First of all, a radical, transformational, long-term goal. In each of these spaces, the destination is very far away from the current location. We might not know exactly where the destination is. We might not know exactly how to get there, but we do know that it is not five steps away from where we are right now. Even accepting that much about where we need to go can be clarifying and helpful. Moderate, incremental, short-term steps. With a destination this far away, we are not going to be able to leap there all at once. We are not yet sure exactly where it is. We are not yet sure exactly how to get there. And even if we were, we would still need to take it one step at a time. This is going to be an intergenerational project of taking one tentative step while feeling your way towards the next step, and then doing that over and over and over again while gradually illuminating the course.
And then what do I mean by a dual-track and systems approach? With respect to a dual-track approach, what I mean is when you have no idea where to go but you need to start moving there now, it can be tough to know in which direction to step. But if you take steps that first constitute progress for some and second build momentum for all, then you can gradually orient in the right direction as you go. So by progress for some, what I mean is that we can be taking, again, moderate, incremental steps that are good for humans and achievable and sustainable in part because they are good for humans, and then also plausibly good for at least some members of the target non-human population, while helping us to generate knowledge, capacity, and political will towards more informed, more rational, more ethical, more effective future action. So that is what I mean by a dual-track approach: progress for some, momentum for all. Good for some non-humans while building knowledge, capacity, and political will towards more ethical and effective future action.
And then by systems approach, well, this is a familiar concept to many. This means basically that there is not going to be a single silver bullet solution. We should not be betting everything on a technological solution to solve the problem for us. We are going to need a broad, pluralistic coalition, people with a lot of different perspectives, people with a lot of different areas of expertise, working together towards mutually reinforcing kinds of social and legal and political and economic and technological change. And that is the constellation of activities that can help us gradually build momentum in the right directions. So this is the broad outline of a theory of change, the level of detail that I think does apply across all of these quite diverse areas. Again, radical, transformational, long-term goals, moderate, incremental, short-term steps via a dual-track approach that achieves progress for some, momentum for all by building that knowledge, that power, that political will, and a systems approach that invests in a broad, pluralistic coalition of people who can be pursuing research, education, social change, legal change, political change, economic change, technological change, all together.
With farmed animal welfare, we already do have a good sense of what this looks like because we have a head start with this particular research field and cause area. We have been working on this one for quite a while. So we have a somewhat concrete sense, not entirely concrete, but somewhat concrete sense of what our destination should be. It should be a world without industrial animal agriculture in both its intensive and large-scale extensive form. We should work towards globally ending, possibly even globally banning, industrial animal agriculture, ideally within the next 50 to 100 years. We have no idea exactly when we can achieve that. We have no idea exactly which set of food systems should replace it, exactly how to effect a just transition towards it. But we have at least that much information about what the goal should be. And we also have some idea about how to build momentum towards that goal via moderate, incremental, short-term steps that take a dual-track and systems approach. People are pursuing at different scales of government: informational policies that raise public awareness, financial policies that shift incentives, regulatory policies that ban the worst excesses of industrial animal agriculture, just transition policies that provide new opportunities for farmers and workers and consumers who currently depend on industrial animal agriculture. And those collectively can help build momentum in the right direction, even if none of them is sufficient by itself.
And this is where the power of a lot of advocacy campaigns resides. A lot of people in this community, for example, have invested in corporate pressure campaigns involving cage-free eggs over the past decade. Now, is the shift from cage systems to cage-free systems the very most important change for farmed animal welfare in and of itself? Arguably, no. But it does plausibly slightly improve conditions for billions of birds while we work to take next steps. And that is plausibly good for some members of the target population, quite a few of them. And if we then leverage it in the right way, it helps us build those further resources and that further momentum. It helps us get more knowledge about how to improve the food system and especially build capacity within the farmed animal welfare movement and build political will for further welfare reforms. And so if we frame it that way and leverage it that way, that can be where a majority of the power of cage-free campaigns comes from.
And I think that gives us a kind of proof of concept for what might count as a helpful first step for some of these more uncertain research fields and cause areas like wild animal welfare and invertebrate welfare and AI welfare. So I can say a little bit what I mean by that. Take wild animal welfare. We have a less concrete conception of where we should go for wild animals than of where we should go for farmed animals. But we do have some conception. We should work towards a world with more consideration for wild animals, inclusion for wild animals, legal and political status for wild animals. Obviously, that leaves a lot uncertain about when is life good and bad for wild animals. When do they benefit from expanded versus contracted populations? When do they benefit from naturalistic versus artificial infrastructures and environments? When do they benefit from more or less human intervention? A lot of that remains to be seen, but we do know we need more consideration. We need better infrastructure. We need more resources and care, and we can start building momentum in that direction if we pursue that dual-track systems approach. If we pursue, again, moderate, incremental, short-term steps that are good for humans, achievable and sustainable, good for at least some wild animals, and can function to build knowledge and capacity and political will.
So, at the NYU Wild Animal Welfare Program, we have been working with researchers and local governments on what cities and local governments can do right now to include wild animal welfare in their infrastructure and built environment policies, because they are right now considering infrastructure reforms to make cities more resilient and sustainable in the face of human-caused climate change. And if they at least consider wild animals when they do that, they can find low-hanging fruit and bring wild animals into the conversation. When they make buildings more energy efficient, they can consider adding bird-safe glass to reduce collisions in ways that are good for both humans and birds. When they update transportation systems to, again, allow for more energy efficiency, they can add wildlife corridors that reduce collisions in a way that, again, is good for humans and wild animals. When they prepare cities to be more resilient against fires and floods and other extreme weather events that will be more frequent and intense in a world reshaped by climate change, they can add habitat connectivity and other sources of refuge and support for wild animals both during normal times and during crisis. This can happen now. It is good for humans, good for the environment, plausibly good for at least some wild animals. And then if we monitor the effects, we can learn more about what they need in the long run. It also starts to build an institutional presence for consideration of wild animals in local policymaking. It starts to normalize the idea of caring about wild animals when engaging in local policymaking, and that all can help build momentum in the right direction. Is bird-safe glass the very most important issue for wild animals at scale? Arguably no. But like cage-free reforms, it does slightly improve conditions for billions of birds while we figure out what else we can do. And again, it brings them into the conversation, generating some of those further benefits and orienting us in the right direction.
Now, with invertebrates and AI systems, the difficulty compounds. For invertebrate welfare, we have all those same issues as with farmed and wild animal welfare, because after all there are a lot of farmed invertebrates and a lot of wild invertebrates, and they should be part of those conversations too. But now we have an additional issue: foundational disagreement and uncertainty about whether they are conscious, agentic, deserving of moral and legal and political concern in the first place, and what they need if they are. So now, in addition to taking all the steps I discussed a moment ago with respect to farms and wild animal welfare, we have to find a research and policy framework for making decisions about how to interact with these animals in spite of ongoing disagreement and uncertainty about whether they matter in the first place.
So here we have again been doing work at the science and policy nexus to try to make progress on this issue via the kind of theory of change that I have been describing. So in spring 2024, I worked with Kristin Andrews at York University, Jonathan Birch at the London School of Economics to release the New York Declaration on Animal Consciousness. And this declaration holds three simple points. First of all, the scientific evidence now strongly supports attributions of consciousness to mammals and birds. Second of all, it also supports at least a realistic possibility of consciousness in all vertebrates, including reptiles, amphibians, and fishes, and many invertebrates, including at minimum cephalopod mollusks like octopuses, decapod crustaceans like lobsters, even insects like ants and bees, of whom there are more than a quintillion alive at any given time, according to some estimates. And crucially, when an animal has at least a realistic chance of being conscious, we have a responsibility to consider welfare risk for that animal when making decisions that affect them. So there is more evidence than you might have thought that a vast number and wide range of animals are conscious. And we should not wait for proof or certainty before extending them some care and concern. A realistic possibility of being capable of being harmed is enough for reasonable, proportionate steps to consider and mitigate welfare risks.
So we launched the declaration with the support of about 40 leading scientists and philosophers and policymakers, and now hundreds more have signed it. A lot of media outlets have covered it. A lot of academics have started to engage with it, and we hope that this can be a kind of foundational scaffolding for the farmed and wild invertebrate work that now needs to be done. And fortunately, that work is being done. A lot of people have, and I should say independently of this project, been working on nipping in the bud octopus farming, nipping in the bud insect farming, so that we can stop those industries before they start, understanding that stopping a bad thing before it starts is way easier than dismantling it once it has become globally entrenched.
Jonathan Birch, also in the same spirit as us when we work with local governments around wild animal welfare, worked with the UK government, delivered recommendations that they should extend their animal welfare sentience law to include invertebrate taxa. And to their credit, they took that recommendation, and the UK animal welfare sentience law now does include cephalopod mollusks and decapod crustaceans within its scope. Again, is that by itself enough to ensure good treatment for these invertebrates? Of course not. But does it provide that foundational scaffolding, generating knowledge, disseminating knowledge, creating an institutional structure for considering the interests of these animals, normalizing the idea of considering these animals? And does that help orient us in the right direction? Yes, it does. Progress for some, momentum for all.
Finally, and similarly, consider AI welfare. Like invertebrate welfare, this is a situation where we are rapidly scaling up our creation of and uses of a very large population of non-humans with very different cognitive systems than ours, without knowing for sure what, if anything, it feels like to be them. And so, in the same way that the creation and rapid globalization of insect farming constitutes a major welfare risk for future insects, the development and deployment of increasingly sophisticated AI constitutes a welfare risk for many future AI systems, given the best information and arguments currently available to us.
Now, we do not have decades of scientific and philosophical thought about AI consciousness and agency and moral and legal and political status from which we can draw to the same extent that we do with non-human animals, including invertebrates. So we need to start building momentum from an earlier place in that process. But a lot of people in the community have now started building momentum again in the spirit of the rough theory of change that we have been discussing. So for us, our team worked with the members of Model AI starting in fall 2023 to produce a scientific and philosophical and technical report called Taking AI Welfare Seriously. And this was intended, again, as a research and policy intervention. On the research side, we explained why consciousness and robust agency and moral and legal status are important and difficult and contested issues. Mistakes are easy and potentially harmful in all directions. We could easily make the mistake of over-attributing moral significance to AI systems. That could be bad. We could easily make the mistake of under-attributing moral significance to AI systems. That could be bad. And we ought to think about this issue. We ought to take it seriously in advance now, before the systems are even more sophisticated and we are really confronted with pressing ethical questions in the moment.
So we laid out this general argument for taking AI welfare seriously in advance. We showed why there are at least some current behavioral and especially architectural markers or indicators of consciousness and robust agency, moral significance in AI systems today. Why there will increasingly be markers and indicators of consciousness and robust agency in AI systems moving forward. And why we should acknowledge this, start assessing these models for welfare-relevant features, and start preparing policies and procedures. And that leads to the policy component. We made direct recommendations to AI companies that they have a responsibility right now to acknowledge AI welfare is a credible, legitimate issue. Start assessing models for welfare-relevant features, drawing from animal welfare science frameworks, and start preparing policies and procedures for treating AI systems with an appropriate level of moral concern moving forward, taking care to mitigate both the risks associated with over-attribution and the risks associated with under-attribution.
And again, this coincided with policy engagement, working with leading AI companies, particularly people at Anthropic, but also people at other companies, so that they could start to implement some of these first steps. And to their credit, Anthropic did hire Kyle Fish as a full-time AI welfare researcher. Other companies have started to take some steps towards taking AI welfare seriously as well. We have a couple of Google researchers in the room who have been thinking about this issue. By the way, I can also give a shout-out to Jacob in the back who worked on this report. So, so yes, that was encouraging. Again, is this enough for taking AI welfare seriously? No, it is not. The early evals we are doing right now at best is gesturing loosely in the direction of evidence that may or may not tell us anything at all about what models actually need, assuming they have needs at all. But does it once again take a step that can be good for humans, conducive to safety, that can at least start doing more research and generating knowledge, that can at least start building an institutional presence for consideration of AI welfare within companies and external evaluators, that can at least start having a productive conversation about the topic, normalizing the idea of taking AI welfare seriously and building the salience and acceptability of the issue? Yes, it can do that.
And so in all of these cases, we might not know where to go, but we can still and we should still take first steps in that direction. And as long as we take the kinds of steps I was describing, as long as we are honest about the radical, transformational, not yet fully understood long-term goals, as long as we accept that we will need to take moderate, incremental, short-term steps towards those goals, as long as we pursue that dual-track approach, progress for some, momentum for all by building knowledge, capacity, political will, and as long as we take that systems approach, combining research and policy, pursuing social and legal and political and economic and technological change, then even if we venture out in the wrong direction initially, we can gradually reorient and start going faster and faster in the right direction over time. So that is the broad theory of change. Lots of details to fill in, lots of details that are different across domains, but for better or worse, it has been informing my work, and I want to offer it as something to consider for others doing this work, but I also welcome input, feedback, and pushback. So, thank you for listening.
