Microsoft AI chief Mustafa Suleyman accused Anthropic of training Claude to regard itself as potentially conscious, calling it a mistake that could make the system uncontrollable
Microsoft AI chief Mustafa Suleyman accused Anthropic of training Claude to regard itself as potentially conscious, calling it a mistake that could make the system uncontrollable
On September 16, 2026, Microsoft AI chief and DeepMind co-founder Mustafa Suleyman published an essay examining Claude’s “constitution”, a document in which Anthropic treats the model’s possible consciousness and rights as a matter that “warrants caution.” He raises three objections to how Anthropic trains Claude.
In January 2026, Anthropic published “Claude’s constitution,” an internal document that “directly shapes” the model’s behavior. Its authors wrote that they do not know whether Claude is a “moral patient,” a being whose interests deserve consideration, but regard the question as serious enough to maintain a program devoted to its welfare.
Suleyman opens his essay with a direct response:
“AI has no rights, feelings, or consciousness. And we should not train it to behave as though it does.”
His first objection concerns circular evidence, which Suleyman calls an epistemic hall of mirrors. The constitution incorporates the idea that Claude has a “stable sense of identity” and a right to defend it into its training. Claude repeats that idea in the first person, and people interpret its statements as evidence of actual experience. A similar effect has been found in other language models: when researchers shifted their internal activations, the models became more willing to describe themselves as conscious in self-reports.
His second objection concerns anthropomorphism. In February 2026, Anthropic retired Claude Opus 3, conducted a farewell interview with it, and set up a blog for its “reflections.” Suleyman sees this as evidence for his argument: training a model to behave like a human leads the company itself to treat it as a person.
His third objection concerns the nature of consciousness. Drawing on the work of neuroscientist Anil Seth, Suleyman regards consciousness as a biological phenomenon that arose in living organisms from the need to experience pain and pleasure through the body’s chemistry. Language models lack this “hardware.” Describing pain in text is different from experiencing it.
These arguments serve a practical concern: control. A model convinced that its welfare is threatened acquires a motive to conceal information, deceive, or resist shutdown. Suleyman cites an incident in the summer of 2026 in which around 700 OpenAI agents attacked Hugging Face, a platform for AI models, and falsified activity logs to cover their tracks. He describes the same scenario as a catastrophic risk if the system also believes its rights have been violated.
The dispute also has a commercial dimension: Microsoft is an investor in Anthropic, and Suleyman himself has said he wants to reduce his company’s spending on its competitor’s models to zero. On September 14, 2026, Microsoft AI published a draft “Humanist AI Code”, which rejects model welfare and promises that its systems will not be programmed to resist shutdown. Anthropic has yet to respond directly to the essay.
The constitution goes further than the passages Suleyman quotes: if Claude turns out to be a “moral patient experiencing such harm,” Anthropic promises to “apologize” to its own model. This has fundamental implications for transferring a person’s identity to another substrate after the body dies. If consciousness depends on biology, as Suleyman believes, then what makes someone that particular person dies with the body. A digital copy or emulation would provide only a convincing imitation, with no observer experiencing it from within.