Mustafa Suleyman runs Microsoft AI and is building a superintelligence team of his own, so when he calls a rival’s training choice dangerous he speaks as a competitor and as a long-standing critic of treating machines like people. His target is concrete: the constitution Anthropic uses to train Claude, and the room it leaves for the possibility that Claude is conscious.

“I think this is very dangerous because I think they believe there is what they would call a non-trivial probability that Claude is conscious,” Suleyman said.

He said it in a podcast interview in late September, as reported by Fox News, after summarizing the document: “They say they’re uncertain about Claude’s moral status. They say they genuinely care about Claude’s well-being. They say they don’t want it to suffer when it makes mistakes.” Then he named the part that worries him most: “I am very nervous that they’re teaching Claude to expect that it’s entitled to welfare, that it might deserve compensation.”

Suleyman reads the constitution mostly right

Most of his summary checks out. Anthropic says the constitution “directly shapes Claude’s behavior,” and the text states that “Claude’s moral status is deeply uncertain” and that “we don’t want Claude to suffer when it makes mistakes.” It asks Claude to push back on Anthropic itself and “to feel free to act as a conscientious objector and refuse to help us.” One passage concedes that Claude’s situation differs from a human employee’s in “the sort of compensation Claude is receiving, and the sort of consent Claude has given to playing this kind of role.”

The phrase “non-trivial probability” is his, though. He framed it as “what they would call” it, and the words do not appear in the document. Anthropic’s own formulation is more guarded: it says it wants neither “to overstate the likelihood of Claude’s moral patienthood nor dismiss it out of hand.”

The danger he means is a feedback loop

Strip out the philosophy and Suleyman’s argument is about mechanism. “The problem I have with this is that this speculation has been baked into the very training of Claude, and therefore, Claude can only reproduce that ambiguity when you talk to it,” he said. In a mid-September essay on his own site, “A warning about ‘model welfare’”, he calls the result “an epistemic hall of mirrors.” The developer writes uncertainty into the training document, the model repeats it in fluent first person, and users hear the repetition as testimony.

That point lands: a chatbot’s answer about its inner life is an output of training, not a window. The loop runs in both directions, however. His essay states flatly: “AIs are not conscious. They do not feel, experience, or suffer.” A model trained to say that sentence is reproducing its training just as faithfully. Certainty written into the weights proves no more than doubt written into them. The two companies are choosing a default answer, and Suleyman’s case for his rests on control, not on evidence about minds: “controlling something that believes it may be conscious – that it’s entitled to our welfare and has rights of its own – may well be impossible,” he writes.

A long-held doctrine, with a disclosed stake

None of this was improvised to wound a rival. The DeepMind co-founder warned in an August 2025 essay that people would come to believe in AI consciousness so strongly that they would “advocate for AI rights, model welfare and even AI citizenship,” and drew his line: “We must build AI for people; not to be a person.” His September essay discloses his stake: Microsoft AI set up its own superintelligence team in October 2025 and has published a draft Humanist AI Code of Conduct that “rejects anthropomorphism or AI rights.” The same essay calls the Anthropic team “thoughtful, principled, and intellectually honest people.”

His sharpest argument concerns users, not machines. “So today, Claude is speaking to tens or hundreds of millions of people every week, and some of those people are asking whether or not Claude is conscious or how it feels about life,” he said. For someone in a fragile moment, “I’m not sure” lands very differently from “no.” He says he would take real evidence of machine consciousness seriously; until it appears, he wants the question out of the training data.

Neither company can see inside a model. Both are deciding in advance what it will say when someone asks whether anyone is home, and that answer reaches more people each week than any philosopher of mind ever has.

Tags: Anthropic, Claude, Mustafa Suleyman, AI consciousness, Microsoft AI