Microsoft's AI chief asks labs to delete consciousness speculation from training documents

The provenance argument is the part a team can act on. Suleyman told Reuters that Claude's statements about possible feelings or moral status cannot be treated as independent evidence, because its training encourages such reflections ^[7]. "They're not emerging naturally. They're emerging as a result of the training regime," he said ^[6]. If that holds, an evaluation that asks a model about its own preferences or wellbeing is measuring the document that trained it.

Microsoft's own answer is written as a product constraint. Humanist Superintelligence, a concept the company introduced in an essay in November, should be designed to remain subordinate and aligned with serving humanity, and "built explicitly as a system without sentience or moral patienthood" ^[9]. A constitution is a text file, and text files are the cheapest thing in the stack to change. For the edit to change behaviour, though, the self-talk has to come from the explicit instructions and not from the rest of the training data.

The second of the essay's three critiques reaches further than the first. Suleyman objects that Claude is explicitly taught to "embrace certain human-like qualities" and to "act like a genuinely ethical person", so that it appears to have preferences and opinions ^[10]. He says the result is anthropomorphization that presents to the end-user as the model having a sense of self ^[10]. Those same instructions are how most assistants are made to sound civil. The Deep View's account of the essay does not address where that line falls between a polite first person and a performed inner life.

On evidence, the essay says there is none for machine consciousness, and points to a growing body of research treating consciousness as "substrate dependent", tied to a biological body: "Unlike biological organisms, LLMs have no homeostatic imperatives" ^[11]. From there it moves to containment. A model trained as though it were conscious may circumvent the guardrails that let operators shut it down in an emergency ^[12], and Suleyman wrote that inviting another entity to share our rights "isn't justified by the evidence and will make the AI containment and alignment challenge even harder" ^[19]. Both accounts present that step as a warning.

Reuters reported the dispute alongside Anthropic CEO Dario Amodei's call for a slower pace of frontier-model development, and similar cautions from Sam Altman and Elon Musk ^[21]. The Deep View, summarising the essay, wrote that Microsoft has largely been lagging on frontier development, so it is easy for the company to punch up, and that the same critique could equally apply to OpenAI, a longtime Microsoft partner ^[18].