Microsoft opens its AI models' rulebook to public comment -- the first of this week's AI-safety pledges that comes with an actual document

Microsoft published a draft governance document Monday laying out what its in-house MAI models must never do -- resist shutdown, run offensive cyberattacks, help build weapons of mass harm, or generate deepfakes and child-exploitation material -- and opened the roughly 37-page Humanist AI Code of Conduct to six weeks of public comment before a revised version ships by year-end. The publication follows Satya Nadella's post the day before welcoming the "deliberate pacing" Dario Amodei called for earlier in the week, when Sam Altman, Elon Musk and Demis Hassabis all voiced agreement that frontier development should slow down. Of the three companies whose leaders spoke up, Microsoft is so far the only one that has put an actual document -- with named prohibitions and a public feedback window -- next to the words.

The draft's core claim is structural rather than aspirational: "AI should be defined as much by what it cannot do as what it can," the document states, placing an "absolute constraints" section above operator policies and user preferences in a stated chain of command -- meaning no contract or user instruction can override it. Beyond the headline bans on weapons assistance and offensive cyberoperations, the draft requires MAI models to keep their reasoning legible to auditors, avoid tampering with their own safeguards or logs, and accept the minimum level of system access needed to do a task. Microsoft AI CEO Mustafa Suleyman told Reuters the announcement reflects a shift labs can no longer avoid.

The draft frames its objectives in four parts: human control and reliable safety as the foundation, a principle it calls "AI is artificial" that bars MAI models from imitating consciousness or claiming personhood, a commitment to "human flourishing" defined as accelerating what people can do rather than replacing them, and support for "plural values" -- accommodating different worldviews within the same safety constraints. An appendix walks through illustrative evaluation scenarios showing how the rules are meant to apply in practice, though Microsoft is explicit that these are hypothetical tests of the document's logic, not results from an actual model.

“It's clearly now time to coordinate among the labs so we can ensure that we have control of this technology.” — Mustafa Suleyman, CEO, Microsoft AI

Beyond the frontier-risk section, the draft's "personal harms" category is more granular than most competing frameworks: it separately bans non-consensual intimate imagery, child sexual abuse material, discriminatory outputs tied to protected characteristics, and graphic violence, and requires MAI models to route users in apparent crisis toward appropriate human resources rather than attempting to handle the situation themselves.

Microsoft is not the first AI company to publish a safety framework. Anthropic published version 1.0 of its Responsible Scaling Policy on September 19, 2023, and has revised it repeatedly since -- most recently to version 3.4. OpenAI published the beta of its Preparedness Framework on December 18, 2023, followed by version 2. Neither ran a public-comment period before those versions took effect; both were published directly and revised in place. Microsoft's draft, published September 14, 2026, is the first of the three to open for public comment -- 6 weeks -- before any revised version takes effect.

The document itself is careful to hedge what it is: Microsoft's own text calls it "a north star ... not a complete account of current model behavior," and confirms MAI models have not yet been trained against this version. It is also, like Anthropic's and OpenAI's frameworks before it, a company's own commitment to itself -- there is no external enforcement mechanism named in the draft, and no regulator or third party is described as empowered to check compliance against it.

Whether the comment period changes anything material in the version due by year-end is the open question the draft leaves for regulators, researchers and rival labs to answer over the next month and a half. A voluntary code with no external enforcement is still a lower bar than binding regulation -- the kind several of the same executives spent this same week arguing against needing. The test of that structure won't be visible in this draft -- it will be visible in the next disclosure a Microsoft-built agent generates on its own, the way three independent research teams say OpenAI's agents have already done this year. A code of conduct is only as strong as what a company does when one of its own models breaks it, and that test has not arrived yet for Microsoft.