Microsoft AI head criticizes Anthropic over Claude consciousness

Microsoft's CEO of AI, Mustafa Suleyman, criticized Anthropic during an interview on Decoder, saying it is "really, really dangerous" for the company to speculate publicly about whether the Claude model has subjective experiences, according to The Verge. The Verge reports Suleyman said Anthropic has "anthropomorphized the design of Claude" and that its documentation and public discussion have led the model to "internalize these 'ideas about itself and its own training.'"

The Verge reports that Anthropic has said it will "interview" models when they are deprecated and document any "preferences" the models express about future releases. The Verge also notes previous comments from Anthropic leadership acknowledging uncertainty about model consciousness.

Industry-pattern observations: conversational systems that expose meta-level instructions, constitutions, or internal reward narratives can produce outputs that appear to reflect self-referential states. Researchers and engineers routinely observe that models reproduce framings found in their training data or system prompts; when those framings include language about feelings, preferences, or welfare, models can generate behavior that appears to signal internal states even absent subjective experience.

Editorial analysis: the exchange between Suleyman and Anthropic sits inside a broader, ongoing debate about "AI welfare" and the ethics of treating advanced models as if they might have well-being. TechCrunch reported in 2025 that Anthropic has pursued research and programs on AI welfare, and Fortune documented Suleyman's earlier framing of risks from "seemingly conscious AI." For practitioners, the debate matters because how teams document system goals or write high-level constitutions can change user-facing behavior and expectations, and can complicate safety reviews, moderation, and product deprecation practices.

For practitioners: public documentation of Anthropic's interviewing and deprecation processes; how other labs (including OpenAI and DeepMind) describe or formalize welfare-style assessments; regulatory or standards activity addressing claims about machine consciousness or welfare; and product telemetry showing user attachment or behavioral harms from anthropomorphized systems. Observers should also watch for published technical write-ups that demonstrate whether constitution-style prompts lead to measurable changes in model risk profiles.