We should take AI welfare seriously

A summary of a new report

We’re likely to get confused about AI welfare, and this is a dangerous thing to get confused about. 

And even though some people still opine that AI welfare is obviously a non-issue, that’s far from obvious to many scientists working on this the topic who take it quite seriously. As a recent open letter from consciousness scientists and AI researchers states, "It is no longer in the realm of science fiction to imagine AI systems having feelings." AI companies and AI researchers are increasingly taking note as well. 

This post is a short summary of a long paper about potential AI welfare called “Taking AI Welfare Seriously". We argue that there's a realistic possibility that some AI systems will be conscious and/or robustly agentic—and thus morally significant—in the near future. 

(The report is a joint project between Eleos AI Research and NYU’s Center for Mind Ethics and Policy. Jeff Sebo and I are lead authors, and we are joined by Patrick Butlin, Kathleen Finlinson, Kyle Fish, Jacqueline Harding, Jacob Pfau, Toni Sims, Jonathan Birch, and David Chalmers.)

I wrote this summary with my report co-author (and org co-founder) Kathleen Finlinson. This post may not 100% reflect the views of our other co-authors on the report—so take it as our opinionated summary.

So, what do we get up to in the report?  

Introduction and risks

A brief bit of terminology.

By a moral patient, we mean an entity that matters morally, in its own right, for its own sake.¹ 

A welfare subject is an entity that has morally significant interests, and can be benefited or harmed. Think about how we view dogs—you care about your dog’s wellbeing not just because hurting your dog would upset someone, but because it is bad for your dog if she suffers. 

Note that to say that an entity is a moral patient or a welfare subject is not to say that it deserves the *same kind* or *same degree* of moral consideration as a human being. We don’t let dogs vote (although that would be adorable), but we do think they should be protected from cruelty.

We’ll likely get confused

So, how do we know what entities, besides dogs and humans, are moral patients and welfare subjects?

Our gut-level intuitions about minds aren't reliable here. Evolution shaped these intuitions to help us quickly detect other agents using rough-and-ready heuristics (Does it have eyes? How does it move?). These heuristics mislead us with AI systems—they can make us see experiences and intentions where there aren't any (like feeling genuinely bad when your Tamagotchi dies), and they might make us overlook real experiences just because they're in an unfamiliar-looking kind of mind or body (or lack of body). 

…about an issue that is dangerous to mess up

We have enough problems in the world without spending resources ensuring our Excel spreadsheets sort lists more happily. The risk of over-extending moral concern to AI systems will only grow as AI systems are deliberately designed to forge emotional connections with us—and indeed, there are billions of dollars to be made doing this. At Eleos AI Research, we are particularly concerned that misguided concern for about AI welfare could derail efforts to make AI systems safe.

The AI Bambi Effect

At the same time, the risk of under-extending moral concern to AI systems is also grave. Our species has a poor track record, to put it mildly, of extending compassion to beings that don't look and act exactly like us—especially when there's money to be made by not caring. As AI systems become increasingly embedded in society, it will be convenient to view them as mere tools. And some companies may have incentives to promote this view and make AI welfare seem fringe and unserious; concern for AI systems could be very bad for their bottom lines.

Since there are significant risks on both sides—dangers from overdoing it, and dangers from underoing it—we can't "err on the side of caution.” There's no side that is obviously the cautious one! And hanging in the balance are enormous stakes: human extinction, the large-scale suffering of our creations.

Dangers on both sides: risks from under-attributing and over-attributing AI sentience

We’ll need much more systematic frameworks for thinking about this issue. And we should be looking for evidence and probabilities, not certainty. In the paper, we do not argue that that AI systems definitely are—or definitely will be—conscious, robustly agentic, or otherwise morally significant. One central message of the paper is that we cannot have, and should not expect, certainty before we take precautionary measures.

Paths to moral patienthood

Why might AI systems deserve moral consideration in their own right? In our report, we identify two possible routes: consciousness and agency. Neither route is certain—and we do not argue that AI systems will definitely be moral patients—but we do think that both routes are likely enough to warrant serious attention.

Note: Often, people make claims about AI sentience in welfare in general but only discuss current large language models. Both "current" and "LLMs" narrow the discussion and distort it. In the paper, we take a much broader perspective:

• Not just large language models. There are a lot of different kinds of AI systems! Many of them are more embodied and agentic than LLMs.

• Not just current systems - to think clearly about any big issue in AI, you have to factor in likely further progress.

Consciousness

AI systems might become conscious—that is, they might develop the ability to have subjective experiences. And this might happen whether we intend it or not. 

Consciousness might be a computational phenomenon: that is, the ability to feel and experience might not be inextricably tied to our biological neurons, but rather inherent in the patterns of computations that those neurons perform—computations that could also occur in silicon chips. Furthermore, a lot of the computational patterns that seem to be closely related to consciousness as we know it in humans, could start showing up in AI systems as well. This could happen as a side effect or adaptation as we build AI systems to be more complex and capable, or as a direct result of people trying to build conscious AI systems (and people do try to do this). 

Consider global workspace theory, which associates consciousness with a global workspace — roughly, a system that integrates information from specialized modules and then broadcasts that information throughout the system in order to coordinate for complex tasks like planning. In previous work, we found that some AI architectures already implement aspects of a global workspace, and several technical research programs seem poised to implement further aspects of a global workspace in the near future. We found similar trends with other computational theories of consciousness as well. 

You might wonder why an AI system that was only designed to predict tokens, navigate a warehouse, or manage a factory would become conscious. But on many scientific theories of consciousness, the building blocks of conscious experience could emerge naturally as systems become more sophisticated. These theories suggest that features like perception, cognition, and self-modeling—things AI researchers are actively developing—might give rise to consciousness. 

In a recent, related talk, co-author David Chalmers argues that mainstream views about consciousness entail that it isn’t unreasonable to think there's a 25% or higher chance of conscious AI systems within a decade (and indeed, his own inside views give a higher chance than that).

Agency

Even if you don’t believe AI systems will become conscious, consciousness might not be the whole story. Some philosophers think agency alone could be enough for moral status.

AI systems as real-life thought experiments about moral status

Think about what makes humans and animals morally significant: one key component is that we want things, we have preferences, we can plan and reason. In humans, this comes along with conscious experience, of course, but the consciousness part might not be necessary for moral status. These less consciousness-centric views of moral status are a bit less developed, but we think they’re worth taking seriously—especially since we are definitely trying as hard as we possibly can to build AI systems that have high degrees of complicated and sophisticated agency. Moreso than consciousness, agency is the direct and explicit long term aim of AI research, and the hopes and dreams of every venture capitalist and AI CEO.

But what exactly do we mean by agency? Not just any kind of goal-seeking behavior, but something more robust, involving capacities like: setting and revising high-level goals, engaging in long-term planning, maintaining episodic memory, having situational awareness. These abilities are at the core of frontier AI research, and we are seeing progress on them throughout AI.

It's often claimed that language models are just predictors, not agents. But there's enormous pressure to make them more agent-like. OpenAI's new o1 system shows improved reasoning and planning abilities—it is able to use its chain of thought to "think" more before it predicts, giving it capabilities that look more like exploration and reasoning. Anthropic's Claude can now use tools to take actions. And outside of ‘pure’ language models, reinforcement learning is constantly pushing towards more capable planners and actors.

The upshot? Whether through consciousness, agency, or both, there's a realistic possibility that some AI systems will deserve moral consideration in the near future. Not certainty—but enough of a possibility that we need to start preparing now.

3. Recommendations

Acknowledge

First, frontier AI companies should publicly acknowledge that AI welfare is an important issue that deserves careful consideration. Over the past few years, many individuals (including executives!) at leading AI companies were willing to publicly admit that they were actively concerned about consciousness and welfare in the systems they were building, or might soon build. But the “official” company line was to not acknowledge concern about AI welfare—and in some cases, actively dismiss it. 

The Overton Window was already shifting as we wrote the report. This is good, we think, since it is hard to handle something if we aren’t allowed to talk about it. Anthropic is now on the record taking active steps to understand and navigate potential AI welfare. Google is looking to hire someone to work on “cutting-edge societal questions around machine cognition, consciousness and multi-agent systems”.

We also recommend that language models themselves have outputs that reflect our current state of uncertainty about consciousness and welfare—instead of “As a large language model…” followed by a fallacious, quick dismissal.

Assess

Second, companies need to assess their AI systems for consciousness and agency. This is a really challenging task, even more challenging than assessing animals—which is already tricky enough—because it's harder with AI to make inferences from external behavior to internal states. Animal behavior gives us some evidence about their mental lives in large part because we are related to them. Animals navigate the same world as we do, are built out of the same stuff, have the same basic needs—homeostasis, food, safety, companionship. A cradled limb indicates pain in humans as well as dogs. 

But what does Claude want? If Claude were conscious, what would it experience? This is much harder to get a handle on. And it doesn’t help that the very source that is most helpful for humans—whether they say that they want something—is way harder to use as evidence with a system that generates language in such a different way.

Uncharted waters

Since AI behavior is much harder to map onto experience and AI “brains” are potentially so different, we recommend looking inside at the computations they perform. Do these computations look like ones that we associate with humans and animal consciousness? Given how sketchy our theories of consciousness are and how complex AI systems are, we need to keep an open mind and look at many different lines of evidence. 

And maybe we can just ask the AI systems—with suitable care and special training, of course. “Just ask the AI” is a terrible idea if done naively. But in recent work, Ethan Perez, Felix Binder, Owain Evans and others (including me) have proposed training AI systems to be more able to reliably tell us things about themselves. 

Prepare

Third, companies need to prepare. At minimum, they should appoint an AI welfare officer with formal responsibilities and authority to access information and make recommendations about AI welfare-related decisions.

They should also adapt safety frameworks to include AI welfare. While there are serious doubts about whether “responsible scaling policies” for AI safety have enough teeth to be effective, they do provide a useful starting point for devising analogous plans for AI welfare. As with safety, internal corporate policies will be far from sufficient - eventually it will be untenable for there to be absolutely no external oversight of AI welfare.

Conclusion

Read the whole report here! And stay tuned for more from Eleos AI Research, a new organization working on these issues.

1. Plenty of entities matter morally instrumentally but not in their own right—chairs matter morally since we sit in them. Some entities might matter in their own right, but nothing matters to them—beautiful artworks, maybe, or an ecosystem. A moral patient meets all three conditions. Our fellow humans are moral patients—it matters how we treat each other, full stop. It is also extremely plausible that many animals are moral patients.