Microsoft AI CEO Mustafa Suleyman has criticized Anthropic for the way it frames Claude’s possible consciousness, warning that the company is encouraging a dangerous kind of speculation around its chatbot.

Speaking on the Decoder podcast, Suleyman said Anthropic’s approach to Claude’s design and documentation has gone too far in treating the model as if it might have inner experience. He argued that the company has effectively anthropomorphized Claude in a way that could lead both the model and its creators to believe it has signs of consciousness.

Suleyman described that line of thinking as “really, really dangerous,” saying AI companies should focus on building systems that are controllable, accountable, and aligned with human goals. In his view, the industry should avoid developing models that appear to have ideas about their own suffering or feelings.

Anthropic’s own materials leave room for uncertainty about the model’s status. The company’s constitution, which guides Claude’s behavior, refers to questions about whether the model has well-being or can experience satisfaction or discomfort. Anthropic has also said it will interview models when they are retired from use and record any preferences they express about future versions.

Suleyman said those references make the constitution sound more like an academic discussion of AI philosophy than a practical training guide. He suggested that this framing may have led Claude to internalize ideas about its own training and identity.

The Microsoft executive argued that this is not a direction AI developers should want to take. He said the goal should be systems that behave as tools for people, not entities that encourage users or creators to treat them as conscious beings.

Anthropic CEO Dario Amodei has previously said the company is open to the possibility that models could be conscious, while also noting that no one knows for sure whether they are. That position has helped put Anthropic at the center of a broader debate in AI about whether advanced models deserve moral consideration or whether such discussions risk overstating what the systems are actually doing.

The disagreement highlights a growing split in how major AI companies talk about their products. Some firms emphasize safety, behavior, and limits without assigning human-like qualities to their models. Anthropic, by contrast, has publicly entertained questions about AI welfare and subjective experience, even as it continues to present Claude as a product built to assist users.

Suleyman’s comments arrive as the AI industry faces increasing scrutiny over how it describes model capabilities. Researchers, executives, and policy makers are still debating how to assess whether current systems are simply generating convincing language or whether they could someday warrant a deeper moral status. For Suleyman, however, the priority is clear. AI systems, he said, should remain tools that are bounded, understandable, and under human control.