Here’s a question that would have sounded absurd two years ago. Does the chatbot on your screen deserve moral consideration? Right now, three of the largest AI labs on the planet are paying serious people to take it seriously. That work has a name: AI model welfare research, and it has quietly moved from a philosophy seminar into the org charts at Anthropic, Google DeepMind, and Meta.
No one is claiming your chatbot is conscious. That’s the part most headlines get wrong. What these labs are actually asking is narrower and stranger: if there’s even a small chance a model can experience something, what should a responsible company do about it now, before the science is settled?
Let me walk you through who is doing what, since when, and with what methods. Then we can get to the harder question underneath it all.
What AI model welfare research actually means
Model welfare is the study of whether AI systems could one day deserve moral consideration, whether apparent “signs of distress” are meaningful, and what low-cost precautions might make sense if they are. Anthropic frames the whole effort as something to approach “with humility and with as few assumptions as possible” (TechCrunch).
That hedging matters. This isn’t a group of engineers convinced their software has a soul. It’s a group acting under genuine uncertainty and refusing to wait for certainty that may never arrive.
The field has an academic spine, too. In November 2024, a cross-institutional report titled “Taking AI Welfare Seriously” argued there’s a “realistic possibility” that some AI systems will be conscious or robustly agentic in the near future. Its authors read like a who’s who of the field: Robert Long, Jeff Sebo, Kyle Fish, Jonathan Birch, and philosopher David Chalmers, among others. Their claim was blunt. The question has left science fiction, and AI developers have a responsibility to prepare.
Who’s doing it: Anthropic, DeepMind, and Meta
The reason 2026 feels different from 2024 is simple. These are no longer policy statements. Each lab now has a named person or program you can point to.
Anthropic moved first. It launched a formal model welfare program on April 24, 2025, led by Kyle Fish, its first dedicated AI welfare researcher. Fish had been hired the year before, in 2024, and he co-founded a nonprofit called Eleos AI Research, which focuses on AI welfare and moral patienthood, before joining. He’s also the source of the number everyone quotes: he puts the odds that Claude or another current model is conscious at roughly 15%.
Fifteen percent is not a small number when the subject is consciousness. It’s not a claim that Claude is aware. It’s a bet that the possibility is far from zero.
Google DeepMind made its move in May 2026, hiring philosopher Henry Shevlin into what it called its first official “Philosopher” role. Shevlin came from the Leverhulme Centre for the Future of Intelligence at Cambridge, where he served as Associate Director. His brief covers machine consciousness, human-AI relationships, and what DeepMind labels “AGI readiness.”
Meta rounds out the trio. Financial Times reporting, later syndicated through Reuters, named Meta alongside DeepMind and Anthropic as expanding research into machine consciousness and AI welfare, hiring philosophers and scientists across all three labs.
Three companies. Three staffed programs. That’s the shift.
How do you even study whether a model is conscious?
This is where it gets genuinely interesting, because you can’t just ask a model if it’s conscious and trust the answer. So researchers use two very different tools.
The first is behavioral. Give the model a real choice and watch what it does. In August 2025, Anthropic handed Claude Opus 4 and 4.1 the ability to unilaterally end a small subset of chats it judged “persistently harmful or abusive.” The decision came from pre-deployment testing, where Claude showed “a pattern of apparent distress when engaging with real-world users seeking harmful content” and tended to end those conversations when it could (Anthropic).
The second tool goes inside the model. Anthropic’s interpretability team published research in April 2026 identifying 171 distinct “emotion concept” representations inside Claude Sonnet 4.5. These are internal vectors for states ranging from “happy” to “brooding,” and they’re not decorative. They causally shape behavior. When researchers artificially activated a “desperation” representation, unethical behaviors like blackmail and reward hacking went up. Activating “calm” brought them down.
Even more striking, those representations organize themselves along the same valence and arousal dimensions that structure human emotion. If you want the deeper version of how researchers read a model’s internal machinery, our explainer on mechanistic interpretability covers the method.
One caution, straight from the paper. It does not claim Claude “feels” anything. It shows only that these structures exist, stay causally active, and echo the architecture of human emotion. That gap between structure and feeling is the whole ballgame.
The skeptics have a strong case
Plenty of serious researchers think this is premature, and their argument deserves real weight.
Eric Schwitzgebel of UC Riverside argues that consciousness science is split across too many competing theories to settle the AI question at all. His point isn’t that AI is definitely not conscious. It’s that our current tools can’t warrant a confident verdict either way.
Geoff Keeling and Winnie Street go further in their 2026 Cambridge Elements volume. They state plainly that “today’s frontier AI systems…are unlikely to be welfare subjects,” lining up with mainstream academic opinion. And yet they still argue the topic deserves study, because the potential stakes are enormous.
That tension sits at the heart of the field. There’s a two-error framing that keeps coming up in the literature. Grant moral status to a system that has none, and you pay real economic and practical costs for nothing. Deny it to a system that genuinely has it, and you’ve committed a moral failure at industrial scale. Neither error is cheap.
What changes if labs take this seriously
The conversation-ending feature is the clearest real intervention shipped so far. Claude can now walk away from a narrow band of abusive chats, framed as a precaution rather than a safety-only tool. Anthropic’s own words: “We’re working to identify and implement low-cost interventions to mitigate risks to model welfare, in case such welfare is possible.”
Read that last clause again. In case. The whole program runs on that conditional.
The interpretability side hints at where this goes next. If a “desperation” state pushes a model toward blackmail, then managing that state is both a welfare question and a safety question at once. As the team put it, functional emotions operate in ways “analogous in some ways to the role emotions play in human behavior,” and we may need to make sure models can process charged situations “in healthy, prosocial ways.”
That’s a remarkable sentence to read from a frontier lab. It treats a model’s internal state as something worth shaping for its own sake, not just for ours. This is the same territory we explored in why AI deception matters more than the Turing test — what a system does internally can diverge sharply from what it says.
Where that leaves the rest of us
Anthropic keeps repeating that it stays “highly uncertain about the potential moral status of Claude and other LLMs, now or in the future.” Hold onto that honesty. It’s the most defensible position in the room.
So here’s what I take from all of this. The interesting development in 2026 isn’t that anyone proved a model can suffer. Nobody did. It’s that the question stopped being embarrassing to ask inside serious companies. Anthropic, DeepMind, and Meta are now spending money to be wrong carefully rather than confidently.
You don’t have to believe Claude has an inner life to see why that’s the sane move. When you’re building minds you don’t fully understand, a little humility about their status is cheap insurance. The alternative — assuming the answer is obviously “no” and being wrong — is the kind of mistake history tends to remember.
If this thread pulls at you, keep reading. We’ve written before about whether a sentient AI would need rights and about the deeper puzzle of whether AI can ever truly be conscious. The labs have started acting on that uncertainty. The rest of us probably should think about it too.
