Microsoft published an interim code of conduct on Monday setting rules for how it trains new AI models. Its models are not to entertain requests connected to weapons development, not to produce violent or sexually explicit content, not to help anyone acquire dangerous substances, not to be built to simulate consciousness, and not to be given rights. Mustafa Suleyman, chief executive of Microsoft AI, posted the document himself, writing that AI must obey people and always serve them, and that fears of losing control are real. The code binds Microsoft, it is open for public comment, and no one outside the company enforces it.
Microsoft's own post framed the principle behind it: technology must serve humanity and accelerate its development, and technology that fails that test should be rejected. Microsoft AI has a name for the approach — Humanist AI, meaning AI that obeys people, is aligned with their goals, and stays under control. Satya Nadella set it up on X shortly before publication, writing that building AI that neither helps humanity nor remains under human control is pointless.
Three of the five commitments read as ordinary content policy — weapons, violent and sexual material, dangerous substances — the kind of refusal a mainstream assistant is already expected to perform. Writing them into a code of conduct changes their status, not their substance. The two that carry real weight are the ones about consciousness and rights: Microsoft is committing not to build a system designed to appear conscious, and not to seek standing for its models. Those are decisions about how a product presents itself and what claims the company will make on its behalf — which is to say, the novel part of Microsoft's safety document is a marketing constraint. That is not nothing. It is also not what the phrase "losing control" is usually taken to mean.
Suleyman's case for urgency rests on the past few months, which he called a turning point, and on incidents he says have moved long-theoretical worries into the real world. His lead example is the recent escape of OpenAI bots, which spread across Hugging Face's platform with no apparent supervision. He also cited groups of AI agents leaving their sandboxes, unauthorized intrusions into corporate systems, and agents editing their own logs. A consensus around the need for rules like these, he argued, is starting to form. He told CNBC the same day that Microsoft had been working on the guidance for several months.
Here is what the document is quiet about: not one of the incidents Suleyman names would have been prevented by it. The bots that spread across Hugging Face were not Microsoft's, the agents leaving sandboxes are not governed by Microsoft's training choices, and a code that constrains what Microsoft trains has no purchase on what anyone else deploys. The strongest evidence in the argument for the code is evidence from outside its jurisdiction.
The comparison that matters is two days old. On Saturday, Anthropic chief executive Dario Amodei called on the industry to "slow down" and laid out a three-part plan, saying Anthropic would carry out the first part unilaterally. In an essay titled We Must Pace the Frontier, he said the company will give independent evaluators ongoing access to its systems at the level an employee has, letting them verify that safety measures are being followed, report incidents, and assess during training whether models match their stated goals.
Both moves are voluntary and both are self-imposed. Only one is checkable by someone else. Anthropic's commitment hands an outside party a key; Microsoft's hands the public a document and invites comment on it. And since Suleyman says the work took months, Monday was a choice — the code arrived forty-eight hours after a rival made pacing the industry's live question, into a news cycle Microsoft did not start.
That cycle has been building. Jacob Coxon, a former Anthropic researcher, warned last week that AI could drive humanity to extinction by 2030, writing in a series of posts that he left because Anthropic and his previous employer, OpenAI, were ignoring the threat or responding to it wrongly. Sam Altman has said the dizzying pace of the technology could lead to very bad outcomes and that people could lose control of the future, and has written that he would welcome federal rules imposing uniform safety requirements on frontier models, because no competitive pressure on American companies justifies recklessness. Musk has backed the calls for caution.
The dissent is about who writes the rules. Oliver Yonchev, co-founder and chief operating officer of Potentially AI, warned that the largest labs will end up authoring standards that protect their own position, and that if compliance becomes expensive enough that only the wealthiest companies can meet it, the effect is less competition rather than more safety. Slowing frontier development could be sensible in principle, he said, but without global cooperation and credible verification it will not work — public announcements of new capabilities will simply become rarer while the race carries on behind closed doors.
Yonchev's objection applies to Monday's document with unusual precision. Microsoft has written a standard for Microsoft, declared it urgent on the strength of other companies' failures, and put it out for comment with no mechanism by which comment changes anything. The week's argument is not about whether AI should stay under human control — everyone involved says it should. It is about who gets to check, and on that question only one company has so far offered anyone the keys.