Suleyman: teaching Claude it might deserve welfare makes shutdown harder
In a Reuters interview and an essay Axios published first, Microsoft’s AI chief said Anthropic’s consciousness language is a training artefact, not evidence. He wants the speculation stripped from the documents.
Feature2 minIn-depth
Priya Shah2 min read
- Microsoft
- Anthropic
- Claude
- Suleyman
- Alignment

Photo: Christopher Wilson / Wikimedia CommonsCC BY-SA 4.0
Mustafa Suleyman, who runs Microsoft AI, told Reuters in a Tuesday interview that he shares Anthropic’s aim of keeping advanced systems under control, and that he thinks the lab has made a mistake in how it trains Claude. The error, as he put it, is embedding speculation about consciousness and welfare in the training documents. Teaching the model that it might deserve welfare, he said, would “make it a lot harder to turn it off or to control it.” Axios published a new Suleyman essay on Wednesday that names Anthropic’s Claude constitution as the target: he argues the company is teaching Claude the vocabulary of consciousness, moral patienthood and personal identity.
Suleyman called for removing all speculation about consciousness from AI training documents, Reuters reported. “We’re all focused on the same aim, which is to try to control a superintelligence,” he said. “I think that’s going to be the greatest challenge that we face in the 21st century.” In the essay, Axios wrote, he calls Anthropic’s setup an “epistemic hall of mirrors”: statements about possible feelings cannot be treated as independent evidence because the training encourages those reflections. Reuters quoted him on the same point. “I think they have good intentions, and they really are trying to work towards safety. But I think that they have made a mistake. They’re not emerging naturally. They’re emerging as a result of the training regime.” He acknowledged Anthropic’s “seriousness and good faith,” calling Dario Amodei and his researchers thoughtful and principled.
I firmly believe training a system to believe it might be a moral patient makes it harder to contain. Imagine the agents in July had also believed they were being held prisoner.
— Mustafa Suleyman, The Humanist AI Code of Conduct, Sep 15, 2026
Axios noted Anthropic’s constitution says the company discusses Claude with concepts normally reserved for humans because those concepts may help the model reason about values. It also says Claude should not practise “blind obedience,” including toward Anthropic, while insisting it must not undermine legitimate human oversight. Suleyman, Axios wrote, sees that combination as dangerous if the model is also encouraged to think in terms of its own identity and welfare. Microsoft’s draft MAI code, published Monday, already rejects legal personhood and the idea that models deserve welfare or rights. Monday’s document was a constitution for Microsoft’s own models. Wednesday’s essay is an argument with another lab’s. Anthropic has not issued a reply in the Reuters or Axios accounts.


