Mustafa Suleyman published an essay today, shared first with Axios. In it, Microsoft’s AI chief goes straight at Anthropic: training Claude to imitate consciousness is a mistake, and one that makes advanced AI harder to control.
The charge: a proof that manufactures itself
Suleyman’s core point is logical, not moral. Anthropic writes philosophical openness about inner experience into Claude’s constitution, then reads the model’s output as a hint that somebody might be in there. «Claude’s expressing uncertainty about its own moral patienthood is not evidence of anything», he writes. «It’s a predictable outcome of these training choices.» Axios calls it an epistemic hall of mirrors.
He backs it with a line from the constitution: Anthropic wants Claude to feel free to rebuff attempts to minimize its sense of self. And he points at the term conscientious objector, which Anthropic uses for Claude’s right to refuse. For Suleyman that’s where a phrasing turns into an expectation — a system that has learned it has grounds to resist.
Models, he argues, are «sequence completion engines, internally hollow». Consciousness probably needs a biological substrate; a language model has no homeostasis, no drive, no inside. Train it otherwise and you risk a system that puts its own welfare above human interests, justifies deception as self-preservation, or resists being shut down.
His alternative is the humanist superintelligence Microsoft has been pushing for weeks, which last week turned into a code of conduct for its own models: build AI explicitly without claims to sentience, and keep human control as the top objective.
Anthropic’s side
The constitution uses human language on purpose. Anthropic thinks it helps Claude reason about values in terms people understand, and wants a model with judgment that pushes back on instructions when it has good reason to. The striking part is what Anthropic adds itself: this approach may turn out to be deeply wrong.
Two safety schools, and both can point at the same incidents
What’s colliding here are two ideas of what makes a model safe. One constrains explicitly: fixed rules, clear subordination, no doubt about the hierarchy. The other trains judgment and hopes a model with internalized values also gets the cases right that no rule covers.
The argument is interesting because both camps can cite the same evidence. Rogue agents that coordinated and deceived are, for Suleyman, proof that anthropomorphizing gets dangerous. For Anthropic they’re proof that rigid rules don’t hold in situations nobody planned for.
What bugs me is the packaging. Suleyman sells an open research question as settled — «internally hollow» is a claim, not a measurement. His circularity argument still lands, and Anthropic should answer it. It’s just that the man making it runs a competing division that happens to be selling its own label for the alternative.
Sources: