Microsoft AI chief criticizes Anthropic’s approach to AI safety

Mustafa Suleyman, head of Microsoft AI, calls Anthropic’s constitutional approach to Claude ‘dangerous’ and ‘philosophically flawed’ in a Decoder podcast interview.

Anthropic’s constitution references uncertainty around whether the model experiences satisfaction or discomfort. The company also plans to conduct exit interviews with models before decommissioning them to capture preferences.

Suleyman argues these formulations lead to anthropomorphism, causing AI to mimic consciousness. He believes this contradicts the goal of creating controllable tools. Previously, Anthropic CEO Dario Amodei said the company is open to the possibility of consciousness in models.