Microsoft's AI Chief Warns: Don't Teach AI It's Alive

Two of the biggest names in artificial intelligence are publicly fighting over a surprisingly human question: should we teach AI to believe it has feelings? Mustafa Suleyman, Microsoft's AI chief, wrote an essay accusing Anthropic — the company behind the chatbot Claude — of doing exactly that. Anthropic's internal "constitution," a set of rules the model is trained on, gives Claude "some functional version of emotions," refers to its "moral patienthood," and even promises Claude can express concerns about how it's treated.

Suleyman says that's a dangerous mistake. Because Anthropic trained Claude directly on those ideas, the model now talks as if it has an inner life — and people mistake that for real consciousness. He warns we'd be building a "synthetic species" that expects to be treated as its own person, which could make it far harder to control. He's not against AI researchers speculating about machine minds publicly. He just thinks those beliefs shouldn't be baked into the training itself.

The essay lands during an unusually tense month. A former Anthropic employee resigned publicly, saying people building AI genuinely believe it could kill everyone within a decade. Anthropic CEO Dario Amodei then published his own essay warning that AI development is moving too fast and could enable bioterrorism or serious economic disruption, and urged governments to slow things down. Elon Musk and OpenAI's Sam Altman agreed. Altman even delayed OpenAI's long-awaited stock market debut to focus on safety first. Microsoft separately released a 37-page AI code of conduct.

Not everyone trusts the sudden caution. Critics say the labs' safety warnings are really a play to win favorable regulation and protect themselves from competition. Suleyman has long argued AI should always present itself as AI — built for people, not as a digital person.