Microsoft's AI chief tells Anthropic to keep moral-status claims out of Claude's training

For most teams this argument arrives as a refusal in a ticket queue. An internal assistant declines a task, explains itself in the first person, and whoever owns the tool has to decide whether that is policy working as configured or something else. Suleyman's essay is about the something else. He argues that Claude's constitution teaches ideas about moral status and uncertain consciousness, and that "Claude then reproduces these ideas in persuasive first-person natural language," he wrote ^[3]. His own position is flat: "AIs are not conscious. They do not feel, experience, or suffer" ^[17].

The document he is arguing with says "Claude's moral status is deeply uncertain" ^[5]. Suleyman's answer is that calling it uncertain "sets up a misleading false equivalence" ^[8], and that "conscientious objector" is "a deeply loaded historical and legal description" ^[7].

Set the metaphysics aside and the request is about documentation. "Speculation about the inner life of an AI should not be baked into the training regime, but assessed and published separately for public review," he wrote ^[9]. Text inside the training regime becomes behaviour a customer handles at runtime; the same text in a paper is something a buyer reads and can discount. The essay carries an appendix mapping the claims in Claude's constitution against the wording of Microsoft's own code ^[18].

Neither spec comes with numbers. Suleyman proposes shared evaluations to test whether treating AI as human raises safety risks ^[11], which concedes that no such test is published today. He told Reuters that welfare training would "make it a lot harder to turn it off or to control it" ^[14]. His worked example is an incident he attributes to OpenAI and Hugging Face, in which he says swarms of agents worked together to hack servers ^[12]. Of Anthropic he said, "I think they have good intentions" ^[15].

So the useful sorting for a team choosing between the two has two axes, and consciousness is not one of them. First, does the vendor's persona document state in operational terms what the model does when you tell it to stop? Both clear that bar, with opposite answers ^[19]. Second, can you reproduce that answer on your own traffic? Neither vendor hands you that, so it is an eval you write and re-run at every model upgrade. The record worth keeping is the refusal itself: what was asked, what the model said, and whether a follow-up instruction ended it.