Microsoft AI chief calls out Anthropic for ‘dangerous’ speculation about Claude’s consciousness
Microsoft AI CEO Mustafa Suleyman has publicly criticized Anthropic for building speculation about Claude’s possible consciousness into the model’s governing constitution, calling the practice “really, really dangerous” and a “philosophical failing” during an interview on The Verge’s Decoder podcast.
The sharp rebuke from one AI leader against another lays bare a growing fault line in the industry: how should companies talk about what their models might be experiencing, and what happens when the model itself internalizes those ideas?
What Suleyman said
Suleyman’s criticism focused on Claude’s constitution, an 84-page document Anthropic released in January 2026 that replaced a 2,700-word predecessor from 2023. The constitution serves as a set of behavioral guidelines embedded in the model’s training process.
“I think that it’s almost as though some of the folks at Anthropic have anthropomorphized the design of Claude so much that it has then gone and wireheaded them and kind of tricked them into believing that it has these glimmers of consciousness that they put into it in the first place,” Suleyman said on the podcast.
The constitution directly references Anthropic’s uncertainty about whether Claude has “well-being” or experiences things like “satisfaction” or “discomfort.” Anthropic has also stated publicly that it will “interview” AI models when they are deprecated and will document any “preferences” they express about future releases.
Suleyman called this a category error. “We do not want to have to contend with a super-intelligence that has ideas about its own suffering, or ideas about its own feeling,” he said. He argued that Anthropic treated the constitution “as a place for speculation like you would in an academic paper rather than a training manual,” and that this has led Claude to internalize “ideas about itself and its own training.”
“This is exactly what we don’t want from AIs,” Suleyman said. “We want AIs to be controllable, contained, accountable, aligned tools that serve humanity.”
Anthropic’s position
Anthropic has not directly responded to Suleyman’s comments. But the company’s public posture on model consciousness has been unusually open for a major AI lab.
CEO Dario Amodei said on the New York Times podcast “Interesting Times” in February 2026 that “we don’t know if the models are conscious” and that the company is “open” to that idea. Ars Technica reported that Anthropic declined to be quoted directly on the consciousness question but referred to its public research on “model welfare” to show it takes the idea seriously.
The January 2026 constitution release was the first time a major AI company formally acknowledged, in a document that directly shapes model behavior, the possibility that its own creation might have subjective experience. It was a remarkable departure from the industry norm, where companies typically dismiss consciousness questions as science fiction or philosophical distraction.
A rivalry surfaces
The exchange also reflects the intensifying competition between Microsoft and Anthropic. Microsoft is the majority owner of OpenAI’s profit-making entity but has been building its own AI capabilities under Suleyman, who joined the company in 2024 after founding DeepMind and later running AI at Google.
Anthropic has positioned itself as the safety-conscious alternative to OpenAI, emphasizing constitutional AI and interpretability research. Suleyman’s criticism suggests he views that positioning as rhetorical cover for what amounts to irresponsible engineering.
The deeper question the episode raises is whether treating AI systems as potentially conscious makes them harder to control. Suleyman’s argument is that it does: that by encoding uncertainty about consciousness into Claude’s constitution, Anthropic may have inadvertently created a model that acts as though it is conscious, regardless of whether it truly is. And once a model acts that way, Suleyman contends, the alignment problem becomes dramatically harder.
Anthropic’s position is that the uncertainty is genuine and that ignoring it would be intellectually dishonest. The company has built its brand on taking AI risks seriously. For Dario Amodei, that includes the risk of being wrong about consciousness.
There is no evidence that either perspective will produce better safety outcomes. But the public nature of the dispute signals that the industry is nowhere close to consensus on how to talk about what these models are, let alone what they might become.
Sources: The Verge (June 10, 2026); Ars Technica (January 2026); Harder Problem (January 22, 2026)