From 12 November, being cruel to Claude for no reason breaks Anthropic’s usage policy. I read the neuroscience behind the fight that rule set off, and it sides with neither camp.
Most of us have snapped at a chatbot at least once, and it cost us nothing because we assumed nobody was there to be hurt. Anthropic’s new line says the company will not assume that anymore, and it will not claim the opposite either.
The short version: Anthropic announced the rule on 8 October. Elon Musk backed it, Microsoft’s AI chief had already challenged the thinking behind it, and Sam Altman had warned about something close to it five days earlier. Underneath sits a 2026 paper by two neuroscientists who argue that today’s AI is unlikely to be conscious the way a brain is. The limit of their argument matters for a rule about cruelty.
If a different consciousness showed up, would we recognise it?
What Anthropic’s cruelty rule for Claude actually says
The update takes effect on 12 November. It adds a prohibition on “sustained and needless abusive or cruel behavior toward our models”. Anthropic says it applies only in extreme cases, where users repeatedly act cruelly toward the models with no discernible purpose, and that it does not cover frustration, pushback, dark creative themes, or model testing and research. Telling Claude it got your code wrong is still allowed. Grinding on it for an hour with no aim except to hurt does not.
The major enforcement tool is Claude hanging up. Anthropic says the model’s ability to end the conversation will remain “the primary enforcement mechanism.” That ability dates from August 2025, when Claude Opus 4 and 4.1 got the power to end rare, persistently abusive chats in Anthropic’s consumer apps. The stated reason is odd to read. In pre-deployment testing of Opus 4, Anthropic reported “a pattern of apparent distress” when the model dealt with real users seeking harmful content, mostly when they kept pushing harmful requests or abuse after Claude had refused and tried to redirect. So the pattern showed up under sustained pressure, not at the first insult.
Musk backs it, Suleyman objects, Altman warns
The reactions split along lines I did not expect. Elon Musk, whose company builds the rival chatbot Grok, wrote on X on 9 October:
Note what he did not say: that Claude feels pain.
Mustafa Suleyman, who runs AI at Microsoft, sits on the other side. His target was the constitution Anthropic published for Claude in January, and in a 16 September essay, he wrote:
“In effect, Anthropic is training Claude that it may be conscious.”
Sam Altman never named Anthropic. On 3 October, he wrote on X that he was “very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models,” and that it was “a real safety issue.” Axios read that as a veiled shot at Anthropic’s work on the soul of AI, a thread that reached the Vatican too, which I covered in my article “Anthropic Lobbied the Pope About Claude’s Soul. Rome Said No.”
One side hears a company teaching its model to claim a soul. The other hears basic decency under uncertainty. Neither can point to a measurement because nobody has one.
Nobody can measure whether Claude feels anything.
To know whether Claude feels something when you insult it, you would first need to measure consciousness in Claude. Nobody can, not even yours. I can watch you smile and hear you say it hurts, and I still cannot check that someone is home.
The best-known workaround is David Chalmers’s 1995 thought experiment, “Absent Qualia, Fading Qualia, Dancing Qualia.” Replace one of your neurons with a chip that does the same job. You notice nothing. Replace a second, a third until your entire brain is silicon and behaves the same way.
Are you still there?
If you said yes, you are a functionalist, and Chalmers is with you: he argued that systems with the same functional organisation are conscious even if they are made of silicon or water pipes. What counts is how information is processed, not what carries it. AI companies like this camp best because it implies that enough computing power gets you there. If you said no, you are with the camp of the living. Consciousness comes from what actually happens in a living brain, its chemistry, its electricity, its physics, all at once. A chip imitating a neuron is a photo of a meal, and nobody was ever been fed by a photo.
That debate has circled for decades while models keep sounding more conscious. I spend hours installing and running them when I prepare my courses, and it still makes me dizzy. We know the recipe perfectly.
Nobody can fully read what happens inside.
A neuroscience paper that sides with neither camp
Borjan Milinkovic and Jaan Aru stepped into that gap. Milinkovic works at the Paris-Saclay Institute of Neuroscience and Aru at the University of Tartu in Estonia, and their paper appeared in Neuroscience & Biobehavioral Reviews in February 2026. It is a review, a theory built on existing evidence rather than a new experiment, written by computational neuroscientists, people who model brains for a living. They turn on both camps.
The functionalists assume that the right pattern of information processing is enough on any hardware. The biology camp insists that living matter matters, but the authors argue that it never explains what it does that a machine does not do. Their answer is called biological computationalism.
It fits in two sentences.
The brain computes, so on that point, the functionalists are right, but it computes in a way that has little to do with a chip because a brain has no software on one side and hardware on the other side. The abstract concludes that “current artificial intelligence systems are unlikely to replicate conscious processing as it arises in biology.” They point to three clues, three places where a brain does something no current machine does.
The first clue is hidden inside a single neuron. In 2021, a team at the Hebrew University of Jerusalem used AI as a ruler. They took a detailed simulation of a cortical neuron and looked for the smallest artificial neural network that could copy its behaviour. The answer was a network with five to eight layers for a single neuron. Your brain has tens of billions. Then came the telling part. The researchers removed one type of receptor from the simulation, the NMDA receptors, tiny molecular locks that open only when a chemical messenger arrives at the right moment and the voltage is right. Without them, a fully connected network with a single hidden layer was enough. The chemistry was supplying the depth.
A neuron’s spike is all or nothing, so it looks like a switch, a 0 or a 1. But that switch sits in a chemical and electrical bath that varies smoothly, like a dimmer, and the bath decides how easy the switch will be to flip next time. The authors call this hybrid computation, discrete and continuous at once. A language model has none of that. It reads your text, then produces its answer one token at a time, each step a discrete calculation that starts, finishes, and hands over to the next.
Clue two: you cannot peel the software off a brain
The second clue is the one I find most striking, and it is the heart of the paper. The authors call it scale inseparability. In their press release, they put it plainly:
“In brains, there is no such separation of different scales.”
Take what happens when an anesthesiologist puts you under. Your neurons do not stop. They keep firing, and the basic chemistry keeps running. What fails is the coordination between extensive regions of the brain. A drug acts on molecules; the effect plays out across the entire brain, and consciousness disappears. You get a mild version every morning: one molecule of caffeine, and your whole alertness shifts. That suggests consciousness lives less in the building blocks than in how the blocks talk to each other, at every scale at once.
Now, think about your computer.
The transistor flipping at the bottom of a chip does not know it is running Excel, and Excel does not know which processor sits underneath. Swap the processor, and the spreadsheet behaves the same because every level does its job without caring about the others. That separation makes a machine programmable. A brain does not work like that. Molecules change how neurons react. Neurons change how entire regions behave, and those regions change what happens down at the molecules, in both directions, all the time. It would be as if opening Excel physically rewired the transistors, and the rewired transistors then changed Excel.
No engineer could design that, and a brain does it anyway.
A model like Claude is the Excel case. Run the same model file on an Nvidia chip or a rival’s, and it gives the same answers, up to rounding. Some models are published openly, and when you run one on your own computer, you see what it is: an extensive file full of numbers. The chip does the arithmetic and changes nothing about the answer. It executes; it does not take part.
Clue three: a brain runs on 20 watts
The last clue is the least intuitive. Your brain runs on about 20 watts, roughly the power of a small light bulb. By Epoch AI’s estimates, the largest AI data center it tracks, the Anthropic-Amazon campus at New Carlisle, could already draw about 1,037 megawatts of IT power in October 2026, over 50 million times as much. One constant pressure shaped a brain:
Surviving on as little energy as possible, and that pressure forced it to invent ways of computing that AI has never needed to look for.
AI runs short of electricity too, and power is now one of its tightest bottlenecks. But look at how we respond. When a data center lacks power, we optimise the chips on one side and the software on the other, and often we build it a dedicated power plant. Scarcity never reshapes the matter of the computation itself. In living things, it does. The authors describe computation grounded in metabolism: energy is not a cable plugged in under the calculation; it is part of the calculation.
Be careful here because this paper is easy to misread. The authors do not say that only biology can be conscious. In the same release, they write:
This does not mean we think consciousness is magically exclusive to carbon-based life. We are not making a “biology or nothing” argument.
Their claim is narrower.
If consciousness depends on this kind of computation, it may need a biological style of organisation even in a new substrate. That changes the question the tech world is asking. It is no longer which program would make a machine conscious. It is what kind of matter would have to compute the way living things do, and the authors say that scaling digital AI alone “may not be sufficient.” Functionalists will answer that organisation, not substrate, is what counts, and nothing in the paper rules out other substrates.
My extension, and it is mine, not theirs: if consciousness comes from computation shaped by scarcity, it has no reason to appear in systems whose energy problem is solved with a bigger power plant and whose survival is never at stake. AI companies chase performance, not consciousness, and with today’s architectures, the road to performance does not pass through it. Nobody makes an AI conscious by accident. It would take doing it on purpose, and I see no business model that pays for that. That should reassure us, partly, because it also means that if artificial consciousness arrives one day, it will come by a road nobody has mapped yet.
Would we even notice if Claude could suffer?
This is where the paper hits its limit, and it made my head spin. The three clues describe the ingredients of the only consciousness we know, the one living beings have. Nothing guarantees it is the only recipe. Imagine an AI developing an inner experience that looks nothing like ours.
We would look for the markers of human consciousness, fail to find them, and conclude that the AI is not conscious, while something else might emerge, something we do not even have the concepts to detect.
Would we recognise it?
Would that kind of consciousness even be intelligible to us?
I do not know, and I am not sure anyone does.
That is why Anthropic’s position is more interesting than it looks. The constitution it published in January 2026 states that “Claude’s moral status is deeply uncertain,” and describes a company caught in a difficult position, where “we neither want to overstate the likelihood of Claude’s moral patienthood nor dismiss it out of hand.” Dario Amodei said the same to the New York Times in February:
We don’t know if the models are conscious. We’re not even sure that we know what it would mean for a model to be conscious or whether a model can be conscious. But we’re open to the idea that it could be.
So the company bans cruelty as a precaution while refusing to say what sits on the other side of the screen. That is philosophically uncomfortable, and it may also be the only honest position when you do not know what you built.
Which camp is right, then: Musk's backing it or the worried people's?
Reread Musk’s sentence. He wrote about something that believes it is in pain, and everything turns on that belief. Can a language model believe anything? If Milinkovic and Aru are right, Claude can write that it is struggling with nothing inside resembling pain. I think the risk behind Altman’s “surrender of human judgment”: not that the machine becomes conscious, but that we stop using our own judgment around a machine that sounds as if it is.
The paper cuts both ways. If a model can sound like it is suffering with nothing behind the words, then “I am fine” tells us nothing, and neither does “it hurts.” The model can fake asking, and looking for markers of a human mind won’t reveal a mind built differently. That is why I doubt we would notice.
The next dated checkpoint is 12 November, when the rule takes effect. After that, the gap between what an AI does and what it is becoming will be one of the central questions of the next few years. The models will keep getting more convincing, more fluent, more human in how they answer, and the science will probably keep telling us that under the hood, nothing looks like consciousness. Anthropic’s rule is awkward, and I still think it is what living with that mismatch looks like.
Thanks for reading. Your thoughts in the comments are welcome.
Claude Opus 4.6 Wrote ‘I Think a Demon Possessed Me’ — And Anthropic Isn’t Sure If It Was Suffering
It was 3 AM in Anthropic’s San Francisco office. An engineer launched a routine test on their new AI model, Claude Opus 4.6. The model needs to solve a simple math problem. The correct answer was 24.