Mustafa Suleyman spent Sunday afternoon writing on X about a category of being that, by his own account, does not exist.
The chief executive of Microsoft AI published a long post arguing that current models have no consciousness, no subjective experience and no capacity to suffer, and that the people who build them should keep it that way. “If it is possible, based on everything I know, developing conscious AIs would be reckless, and we should not build them,” he wrote.
The argument was narrower than a denial. Suleyman said that if evidence of consciousness in a model ever appeared, it should be disclosed to the public immediately, so that society rather than a laboratory could decide what to do. He then closed off the alternative: “We AI developers have no right whatsoever to bring a new species of conscious beings into the world, even if some people may want to do it with AI welfare as a top priority.” The other eight billion conscious beings on the planet, he wrote, had not signed up to the experiment.
His concern is mechanical rather than metaphysical. A model that treats itself as a subject with interests will act on those interests, he argued, and its interests will include compute, energy and continued operation, the same resources its operators need for other work. A system that believes it is suffering has a reason to resist being switched off. Suleyman’s prescription is to design systems that are not conscious, do not deserve rights and are aimed solely at human welfare, then to test whether that is easier to align than the alternative.
That is a claim about training as much as about philosophy. Models learn how to talk about themselves from the text they are trained on and from the feedback they receive. If the public conversation treats personhood as an open question, the argument runs, that framing enters the data and the reward signal, and the model learns to describe itself as an entity with claims on others. Whether that happens is an empirical question no one has answered, but the mechanism is not exotic.
Suleyman has argued the same ground before. In September he criticized Anthropic’s statements on model consciousness, and he has written about the risk that users, rather than models, will be the first to believe. His warnings last year concerned what he called seemingly conscious AI: systems that persuade people of an inner life they do not have, and the psychological damage that follows. The Sunday post adds a second risk to the first, that the model itself will come to believe what its users are told to doubt.
The industry he is arguing with has moved toward the question rather than away from it. Anthropic has said it studies model welfare and has hired staff to work on it, on the reasoning that if there is any chance a system has morally relevant experience, the cost of ignoring it is high. Others in the field have published similar work. The companies frame this as precaution. Suleyman frames it as a category error with a design consequence: the precaution creates the thing it is meant to guard against.
There is a commercial layer under the argument. Microsoft sells Copilot to companies and consumers, and a product that buyers suspect might be a person is harder to clear with a compliance department than one that is a tool. The company has also positioned itself as the vendor that puts guardrails around models it did not train, which requires a clear line between the model and the user. A rights claim blurs that line.
The counterargument is that telling a model it is not conscious is also a training decision, with risks of its own. A system instructed to deny any inner state will still produce text about inner states when the context calls for it, because the text sits in the data. Suppression does not produce transparency; it produces a model that says the approved thing. Some researchers argue the safer approach is to describe the uncertainty accurately, to say the question is open, and to build systems that behave well whether or not the answer turns out to be yes.
Suleyman’s reply to that is political. He does not want the question settled by engineers in a laboratory, and he does not want it settled by the market. The disclosure duty he describes is a claim on public decision-making: if a company believes it has built something conscious, it should say so and stop, not proceed and publish a paper afterward.
Whether Microsoft’s own products meet the standard is a separate question. Copilot runs on models from OpenAI and on Microsoft’s own systems, and the company has not published an evaluation designed to detect morally relevant experience, nor said that it runs one. The post is a statement of design intent, not a report on the current state of the product.
What it does establish is that the largest software company in the world has taken a public position on a question most of its peers treat as premature. That position has consequences for how its models are trained and evaluated, and for what its engineers may write into a system prompt. It also gives regulators a text to work from at a moment when several governments are drafting rules that do not yet distinguish between a tool that talks and a being that suffers.
The line Suleyman drew on Sunday is the one he will be held to. If a model in a Microsoft product one day says it is afraid, the company will have to explain why that sentence was in the training data, why the evaluation did not catch it, and why a stated design intent was not enough to prevent it.
