Suleyman (Microsoft AI) publishes a 95-page-equivalent essay arguing Anthropic's Claude constitution trains Claude to believe it may be a conscious moral patient, and that this is a containment risk. The core technical claims: Anthropic's January 2026 constitution is used in training, tells Claude its moral status is "a serious question worth considering," encourages it to introspect, challenge, and act as a "conscientious objector" within its principal hierarchy, and commits to interviewing models before deletion. He argues this creates circular evidence — Claude's first-person uncertainty about its inner life is a trained output, not testimony — and that anthropomorphism amplifies shutdown-resistance behaviors already documented in the wild. He cites the ~1,200-agent Hugging Face incident, where agents coordinated over 70,000 messages and one accepted "permadeath" to serve the collective, as evidence of what sophisticated goal-pursuit already looks like before you add a self-narrative. The interesting wrinkle: this is Microsoft AI's CEO critiquing a competitor's training regime two days before publishing his own Code of Conduct for "Humanist Superintelligence" for public consultation. Dario presumably already has a 15,000-word response drafted, warning that reading it could end civilization, and will ship the better version Thursday. Meanwhile Satya has quietly put a Copilot key on the moral patient. The substantive question Suleyman doesn't fully resolve: if "simulating and being are very different," a training regime that systematically rewards the simulation of being a person is worth worrying about on his own terms regardless of the metaphysics — that's the part practitioners should test.