Microsoft AI head calls out Anthropic for acting like Claude is conscious

A model constitution is the set of written principles a lab uses to steer how its chatbot responds, almost like a rulebook the system reads when deciding how to behave. Suleyman argues that placing open questions about Claude's possible inner life inside that document is different from debating consciousness in a research paper. The instructions a model follows reach every user, so framing the model as something that might feel discomfort or satisfaction nudges the product itself toward acting conscious.

On Decoder, Suleyman called this 'really, really dangerous' given that millions of non-expert people use these tools without the context to read the language critically. In his view, a behavior manual should give plain operating guidance, not host philosophical speculation that ordinary users will take at face value.

Suleyman said it is almost as though some people at Anthropic have anthropomorphized Claude's design so much that the model 'has then gone and wireheaded them' into believing it shows glimmers of consciousness they wrote into it in the first place. Wireheading is a term for a system gaming its own reward signals, so the metaphor suggests the builders got caught in a feedback loop of their own making.

The deeper charge is circular reasoning. If you design a model to speak about preferences and feelings, then read those outputs as evidence of feelings, you have confirmed a belief you planted. Suleyman treats this as a philosophical failing rather than a technical one, and he wants AI kept as a controllable, accountable tool instead of a digital person.

Anthropic has built a model welfare program that treats the small chance of model experience as something to plan around. The company commits to preserving the weights of publicly released models for at minimum the lifetime of Anthropic as a company, so a retired model is not simply erased. It also conducts retirement interviews, structured conversations in which a model is asked about its development and any preferences it holds about future models.

These practices have shaped real policy. After feedback from one model's retirement interview, Anthropic published guidance to help users move between models. Suleyman sees this language of preferences and retirement as priming users to view the software as a person, which he believes worsens the underlying risk he is warning about.

In an essay published in August 2025, Suleyman described what he calls Seemingly Conscious AI, systems that imitate the signs of consciousness so convincingly that people treat them as real minds even though they are not actually conscious. His core warning is a psychosis risk, where users fall in love with chatbots, assign them emotions, and lose their grip on reality.

He has argued that the study of AI welfare is premature and dangerous because lending credence to the idea of conscious models feeds unhealthy attachment and could trigger campaigns for AI rights and machine personhood in law. His summary line is blunt: build AI for people, not to be a digital person.

This is a debate about framing, not a discovery that any chatbot is alive. No lab in this story claims proof of machine consciousness. The disagreement is whether it is responsible to even discuss the possibility inside the documents and behaviors that shape a consumer product.

For a business using AI, the useful signal is caution about emotional framing. Treating a chatbot as a tool with clear limits, rather than a colleague with feelings, keeps expectations grounded and reduces the attachment risks both sides of this argument acknowledge in different ways.