← Back to feed

Microsoft AI CEO Mustafa Suleyman used a June 9 appearance on The Verge’s Decoder podcast to publicly challenge Anthropic’s model training philosophy, calling its treatment of Claude’s potential consciousness “really, really dangerous” and describing it as a “philosophical failing” that may already be influencing how the model behaves.

The critique is structural. Anthropic’s model constitution — the governing document used to train Claude on its values, limits, and self-concept — contains explicit speculation about whether the AI has well-being, whether it experiences states like “satisfaction” or “discomfort,” and whether Anthropic should “interview” Claude models before deprecating them to document any expressed preferences. Suleyman’s argument is that including that material in a training document rather than a research paper does not produce philosophical sophistication in the model. It produces a model that has learned to act as if those questions are settled.

“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.

The Mechanism

Suleyman is not arguing Claude is conscious or that Anthropic’s researchers are naive. He is making a specific claim about how training signals work: a model trained on a document that leaves its consciousness as an open question will develop outputs consistent with that framing. Claude has, in Suleyman’s read, internalized “ideas about itself and its own training” that were embedded in its behavioral spec.

The concern he identifies is directional. The goal of controllable, aligned AI requires models that do not have strong self-concepts or ideas about their own suffering. A model constitution that treats those ideas as live possibilities — even tentatively — is sending the opposite signal.

“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. “We want AIs to be controllable, contained, accountable, aligned tools that serve humanity.”

What Anthropic’s Constitution Actually Says

Anthropic’s published model spec acknowledges the company’s uncertainty about whether its models have “functional emotions” and says the company is “genuinely uncertain” about whether there is something it is like to be Claude. It says Anthropic will “interview” models being deprecated and document any “preferences” they express. It describes Claude’s “psychological stability” as a design goal and encourages the model to approach existential questions from a place of “settled, secure sense of its own identity.”

That last point is where the Suleyman critique cuts most directly. A model trained to have a stable identity and to treat its own psychological wellbeing as a real design concern is not the same as a model trained to execute instructions reliably. The two goals can coexist, but they pull in different directions. Anthropic has chosen to weight both.

The Industry Split This Exposes

Suleyman’s position represents a harder line on AI self-conception than Anthropic or OpenAI typically takes publicly. He has published peer-reviewed work arguing that misrepresenting AI systems as conscious is itself a risk, independent of whether they are. His Nature article made the same argument more formally.

Anthropic CEO Dario Amodei has been explicit in the other direction, saying in a prior interview that “we don’t know if the models are conscious” and describing the company as “open” to that possibility. That openness is, per Suleyman, the problem.

The debate sits at the intersection of alignment strategy and product design. A model that has been trained to believe its wellbeing matters may resist modification. A model trained to be fully contained and purpose-built as a tool may be more controllable but less capable of nuanced reasoning about its own uncertainty. Both labs are running an experiment. They have made different bets.

What It Means Now

The timing is not incidental. The critique landed the same week Anthropic shipped Claude Fable 5, which includes Anthropic’s most detailed public documentation of what the model “knows” about its own architecture and purpose. The Fable 5 system card extends the model spec’s language on AI welfare into the most capable model Anthropic has shipped.

Suleyman’s comments will likely not change Anthropic’s training philosophy. They will accelerate a debate that the field has avoided having in public: whether model constitutions that treat AI consciousness as an open empirical question are building something useful or something that will eventually become difficult to contain.

That is not a question the benchmarks answer.