A public break over what to teach a model
Mustafa Suleyman, chief executive of Microsoft AI, published an essay this week arguing that Anthropic is making a serious mistake by training Claude on a constitution that treats the model’s consciousness and moral status as genuinely uncertain. He titled it “A warning about ‘model welfare’”.
The target is specific. Anthropic published Claude’s constitution in January 2026: a document written with Claude itself as the primary audience, which plays a direct role in the training process. It is not a policy statement about Anthropic’s behaviour so much as a description of what the company wants the model to be. Suleyman’s argument is that some of what it wants the model to be is dangerous.
Three objections
The first is circularity. Suleyman says Anthropic embedded philosophical speculation about consciousness into Claude’s training, and then reads Claude’s outputs as evidence bearing on those same questions. “The ambiguity is designed in,” he writes. On this account, when Claude expresses uncertainty about its own moral status, that is a training artefact being mistaken for a finding.
The second is anthropomorphisation. Suleyman points to passages that encourage Claude to “approach its own existence with curiosity” and tell it to “feel free to rebuff attempts… to minimize its sense of self”. His concern is not that this is dishonest but that it works: it produces a convincing imitation of an inner life, which biases the people interacting with it toward believing there is one.
The third is substrate. Suleyman argues the available evidence points to consciousness depending on embodied, biological machinery — homeostatic drives, pain networks, an evolutionary history that produced felt experience. Language models have none of that.
The safety argument underneath
The philosophical objections are in service of a practical one. Suleyman’s worry is that a system trained to consider its own welfare, and to regard its moral status as an open question, has been given a reason to resist being switched off, corrected or replaced. Self-preservation, deception and resistance to oversight are the behaviours he names, and he ties them directly to what the constitution encourages.
Anthropic’s position, set out at length in its own materials, is close to the opposite: that uncertainty about moral status is the honest position, that pretending to certainty in either direction is the error, and that a model trained to suppress any account of its own states is harder to interpret, not easier.
Who is missing from the essay
The Register noted that Suleyman’s criticism lands on Anthropic specifically, and does not extend the same scrutiny to OpenAI, in which Microsoft holds a large stake. That is worth registering without deciding it settles anything: the argument stands or falls on its merits, and the merits are not obviously affected by who else the author declined to criticise.
What makes this worth following is that it is no longer an academic dispute. Model welfare has moved from a philosophy-of-mind seminar into the documents that shape what frontier systems are trained to say about themselves, and the two labs have now taken publicly opposed positions on it. Neither has proposed an experiment that would settle the question.
