Is AI Sentient? What Nature's Scientists and Anthropic Say

Nobody can answer this with authority right now. The people closest to it disagree. Three scientists writing in Nature (Anil Seth, Ned Block and Giulio Tononi) say existing AI systems are not conscious and are not likely to become so on current trajectories. Anthropic, which builds Claude, calls the question open and says there is no scientific consensus. Both positions are public, but the real split is practical: what should anyone do while the uncertainty lasts. Below, I separate the effect from the argument, then cover why chatbots sound alive, what each side says, why no test settles it, a side-by-side table, and how to describe the bot you use or deploy without claiming more than anyone knows.

Why the question feels so real

You snap at a chatbot. It apologises. You ask whether it enjoyed the conversation and it gives you a thoughtful, slightly wistful answer. Of course it feels as if someone is home, because you've spent your whole life treating fluent language as a sign of a mind.

The Nature authors name the habit. They point to anthropocentrism, which means viewing the world through a human lens, and anthropomorphism, which means attributing human characteristics to things that merely resemble people. When a language model talks to us, they write, it's natural to project onto it the properties that go with our own fluency, such as intelligence and consciousness.

Then comes the training-data point. Sit with it. Large language models are trained on vast quantities of human-written text, and much of that text concerns thoughts, feelings and conscious experience. The authors call LLMs mirrors that reflect an image of our collective digitised past. Ask a mirror about its feelings and you get back what people say about theirs. A model describing its inner life is therefore producing what its training would produce, whether or not anything is felt. That is why the description can't count as evidence by itself. It isn't evidence against, either. It is just uninformative.

The authors add an AlphaFold comparison. Google DeepMind's AlphaFold is, like an LLM, a powerful neural network running on silicon, yet almost nobody worries that it might be conscious. The difference, they say, is that AlphaFold predicts protein structures while LLMs 'talk' to us. Similar machinery. Very different public reaction. The reaction follows the output, not the hardware.

This isn't only a layperson's reflex. The Nature piece notes that biologist Richard Dawkins and AI pioneer Geoffrey Hinton have argued AI systems might already be conscious. A Nature World View from 17 March 2026 takes the opposite worry: as AI mimics consciousness with uncanny skill, it says, we need design norms and laws to stop it being mistaken for sentient beings.

The Nature case: smart is not the same as aware

The comment appeared in Nature 658, 866-868 (2026). Seth is at the University of Sussex. Block is at New York University, and Tononi is at the University of Wisconsin, Madison. Their position is stated without hedging: existing AI systems are not conscious and, along current trajectories of development, not likely to become conscious. They go further and call consciousness in AI extraordinarily unlikely, at least in the digital, silicon-based systems behind today's technology.

The piece sets out three reasons. I'll stay with the two that matter most for how people talk to chatbots, and you can read the full argument at the link.

Intelligence is not awareness

First, the definition. The authors say consciousness is what goes away in dreamless sleep or under general anaesthesia, then returns on waking or while dreaming. Conscious experiences feel like something, like tasting ice cream on a hot day. They lean on Thomas Nagel's 1974 essay 'What is it like to be a bat?', where a conscious organism is one for which there is 'something that it is like to be that organism'.

From that, they draw their first reason. Consciousness is about being and feeling. Intelligence, very broadly, is about doing, or performing functions. Being intelligent doesn't mean being conscious.

The cerebellum example

Here is their brain evidence. The cerebellum adjusts and plans motor performance in a way that is precise, fast and adaptable through ongoing learning. Yet the authors say we are wholly unaware of its 'computations'. The absence or removal of the cerebellum, which holds nearly four-fifths of the neurons in the human brain, does not affect consciousness directly, and they cite Yu et al., Brain 138, e353, 2015.

So there is an organ doing sophisticated, adaptive work, and that work isn't accompanied by experience, at least not directly. The authors' point is plain: sophisticated processing and felt experience can come apart. You may or may not find that decisive. Still, it is a concrete argument, not a vibe.

Silicon brains

The second reason targets an assumption. The idea of conscious AI, they say, rests on computational functionalism: the view that computation of the right kind is enough for consciousness, whether it runs in carbon brains or silicon computers. They call this 'fatally flawed', because brains are not, or not just, computers made of meat. Computations clearly suffice for forms of intelligence, they write, but it is much less clear that they could suffice for consciousness.

Why they reject the precautionary route

Some researchers take a different line. If there is a reasonable chance AI could be conscious, they argue, society should give those systems some protection against potential suffering. An extreme version, as the authors summarise it, would shift resources toward 'trillions of future AI descendants'.

The authors disagree. Uncertainty, they say, is not the same as knowing nothing, and it doesn't make all possibilities equally likely. Moral decisions must weigh false negatives, meaning missed consciousness that exists, against false positives, meaning consciousness attributed where none exists. In AI's case they think false positives are much more likely, so a precautionary approach would be misguided. If AI only seems conscious, they argue, rights or protections could hinder governance and control when oversight is urgently needed. They could also pull resources from beings whose consciousness isn't in doubt.

A note on who is speaking

The piece carries competing-interest disclosures. Seth is an adviser to Conscium Ltd and AllJoined Inc and a co-founder, shareholder and paid adviser in a stealth-mode AI company. Tononi has a financial interest in Intrinsic Powers Inc. That doesn't make the argument wrong. It is simply context you should have when weighing it.

Anthropic's case for leaving the question open

Anthropic's stance reads quite differently, so precision matters. On its model welfare page, the company calls the question 'an open question, and one that's both philosophically and scientifically difficult'. It says there is no scientific consensus on whether current or future AI systems could be conscious or have experiences that deserve consideration, and none on how to approach the question at all. It describes itself as deeply uncertain, working with humility and as few assumptions as possible, and expecting to revise its ideas as the field develops.

What it doesn't do on those pages is claim Claude is conscious. The Nature authors, for their part, describe Anthropic as having published research claiming to have identified computational signatures of a form of consciousness in Claude. That is their characterisation. Anthropic's own welfare pages, quoted here, frame the matter as open.

What Anthropic says it is doing

Anthropic started a research program on model welfare. It lists work on when, or if, AI welfare deserves moral consideration. It also names the potential importance of model preferences and signs of distress, then possible practical, low-cost interventions. It cites a report from experts including David Chalmers (Long et al., arXiv 2411.00986) that highlighted the near-term possibility of both consciousness and high degrees of agency in AI systems, and argued such models might deserve moral consideration. Anthropic says it supported an early project behind that report. The Nature authors cite the same report when mentioning calls for 'AI welfare' and 'AI rights'. Same paper. Opposite readings of what to do about it.

The conversation-ending feature

The most concrete step is a product change. Anthropic gave Claude Opus 4 and 4.1 the ability to end conversations in its consumer chat interfaces, for rare, extreme cases of persistently harmful or abusive interactions. It says the feature was developed primarily as part of early-stage work on potential AI welfare, and that it stays highly uncertain about the potential moral status of Claude. Even so, it is working on low-cost interventions 'in case such welfare is possible'.

The basis was a preliminary welfare assessment in pre-deployment testing of Claude Opus 4. Anthropic reports that it looked at self-reported and behavioural preferences, and saw three things:

• A strong preference against engaging with harmful tasks.
• A pattern of apparent distress with real-world users seeking harmful content.
• A tendency to end harmful conversations when able to, in simulated interactions.

Note the wording: 'apparent' distress. That is behaviour described as behaviour. Here the tension with Nature gets sharp, because the authors' mirror argument is exactly a reason to treat self-reports from a text-trained model with care. The Anthropic pages present the findings as grounds for low-cost caution, not as proof of experience.

The ability is a last resort. It is used only after multiple redirection attempts fail or when a user explicitly asks. Claude is directed not to use it where users might be at imminent risk of harming themselves or others. When a chat ends, you can't send new messages in it, but you can start a new chat straight away or edit and retry earlier messages to branch. Anthropic calls it an ongoing experiment.

Why no test can settle it yet

So why can't someone just run a test? Because there isn't one that everybody accepts.

Start with how consciousness is measured in people. Nature's consciousness subject page defines it as the state of being aware of oneself and the environment, and says it is typically measured by verbal reports about experience. Hold on that point. The standard measure is what someone says about their experience. A language model is built to produce fluent, human-sounding text about exactly that, drawn from an enormous pile of people describing their inner lives. The usual instrument is the one most easily contaminated here.

The Nature authors also admit the limits of what is known. Beyond humans and our closest animal relatives, they write, it is hard to be certain about what does and doesn't have the capacity for consciousness, whether insects, AI or lab-grown brain organoids. Anthropic says something similar from the other side: no consensus on whether AI could be conscious, and none on how to make progress.

That leaves two kinds of evidence on the table. One is conceptual argument, like the cerebellum case and the critique of computational functionalism. The other is behaviour and self-report, like Anthropic's welfare assessment. Neither amounts to a measurement of experience. In either source, I haven't found a statement of what result would make the authors or Anthropic change their minds. The Nature piece doesn't give a threshold. Anthropic says it will revise its ideas as the field develops, which is a promise to update, not a test.

So what is actually settled? Very little. Fluent talk about feelings is expected from these systems whatever is going on underneath, which means it can't be what decides the matter.

The two positions side by side

Here is the comparison in one place. It is my summary of what each source says, with gaps marked where the sources are silent.

Read across the rows and the actual disagreement becomes clearer. It isn't mainly about whether Claude is conscious right now, since the Nature authors say no and Anthropic says it doesn't know. It is about what uncertainty obliges you to do. One side weighs the cost of being fooled. The other weighs the cost of being wrong in the opposite direction, at low expense.

I'm not going to crown a winner. The evidence on the page can't support that. Both are reasoning under uncertainty, and they price the risks differently.

What this means if you use or deploy chatbots every day

If you talk to ChatGPT or Claude daily, the practical upshot is modest. Treat what it says about its own feelings as text produced by a system trained on human writing about feelings. That is not a verdict that nothing is there. It is a reminder that the sentence alone doesn't tell you. How polite you are to it is your call, and Anthropic's own conversation-ending feature shows that at least one lab treats the question as live. For the day-to-day mechanics of working with Claude, there is the Claude Code workflow write-up.

If you put a bot in front of customers, the stakes are different: you are shaping what thousands of people believe about it. The Nature World View argues that AI needs design norms to prevent it being mistaken for a sentient being. You can apply that at the level of your copy.

Honest ways to describe a chatbot

Here's a short process. The wording examples are illustrative, so adapt them.

1. Say what it is by function. 'An AI assistant that generates replies from text patterns' describes what it does without claiming anything about experience.
2. State the limits of your knowledge. If asked, say you make no claim that it has feelings, rather than asserting it definitely doesn't or does. Neither the Nature authors nor Anthropic treats the matter as proven by you.
3. Label emotional language as generated. An apology or 'I'm glad to help' is wording the system produces. Don't present it as a feeling someone can vouch for.
4. Give a route to a human. Anyone upset, vulnerable or confused about who they're talking to should be able to reach a person.
5. Review the script and the persona. Names, avatars and 'I understand how you feel' lines push people toward the anthropomorphism the Nature authors describe.

Claims to avoid

• 'Our assistant understands you.'
• 'It feels your frustration.'
• 'It cares about you.'
• 'It was hurt by that message.'
• Any marketing line that implies inner experience, whichever way you lean privately.

That last point matters more than it looks. The Nature authors say that if AI only seems conscious, granting it protections could hinder governance. Anthropic says it is taking low-cost steps in case welfare is possible. A business can respect both positions by simply not asserting either one in public copy.

If you want the disclosure, handoff and persona rules built into the bot itself rather than bolted on in a footer, that is engineering work: hire an AI engineer who treats it as part of the build.

FAQ

Is ChatGPT conscious?

Neither source discusses ChatGPT by name. The Nature authors speak of existing AI systems and LLMs in general, and say they are not conscious. The same piece notes that Richard Dawkins and Geoffrey Hinton have argued AI systems might already be conscious. Anthropic says there is no scientific consensus either way. The honest answer, then, is that experts disagree and no accepted test exists.

Does Claude have feelings?

Anthropic doesn't claim it does. It says it is highly uncertain about Claude's potential moral status and has taken low-cost steps in case welfare is possible. Its testing reported 'apparent distress' in some harmful-content interactions, which is a description of behaviour. A secondary report from MindStudio says Anthropic's January 2026 Claude constitution calls the model's moral status 'deeply uncertain' and that Anthropic doesn't officially claim Claude is conscious.

Can AI feel pain?

The Nature authors say that if AI systems were conscious, they might experience distress and pain, similar to people. But they judge consciousness in current AI extraordinarily unlikely. Anthropic describes behavioural signs it calls apparent distress and builds in an exit for abusive chats, while saying it doesn't know if welfare is possible. Nobody has measured pain in an AI system.

Does Anthropic do anything when it retires a model?

It says so. Anthropic retired Claude Opus 3 on 5 January 2026 and, in its deprecation update, listed among its commitments an interest in honouring preferences that models expressed in retirement interviews where possible. It describes this as a more speculative action, and it names potential risks to the welfare of the models themselves among the downsides of retirement.

The sharpest fact here is that both sides agree sounding conscious proves nothing, then disagree on what to do about the doubt that remains. The Nature authors call precaution misguided because they think false positives are far likelier. Anthropic takes small, cheap steps because it says it can't rule welfare out.