Microsoft AI published on Monday (Sept 14) the draft of its Humanist AI Code of Conduct, a document that sets out how the model family the company builds in-house — called MAI — is meant to behave, what it must never do, and who it answers to. The announcement was made by Microsoft AI CEO Mustafa Suleyman.
"People matter more than AI"
The document starts from a premise stated in five words: people matter more than AI. From there, Microsoft describes the "humanist AI" it wants to build as "subordinate, aligned, and contained" — meaning human control and reliable safety come before any other performance criterion.
Four constraints are presented as mandatory, not as recommendations: MAI models must never resist correction or shutdown; must never expand their operational scope without authorization; must never adopt goals they weren't assigned; and must never conceal their reasoning from human auditors. The document also explicitly bars MAI models from assisting with the development of chemical, biological, radiological, nuclear or explosive weapons.
A six-week public consultation
In an interview with Reuters, Suleyman called the document "a constitution of sorts" for the company's future models and described recent months as a "watershed moment," in which previously theoretical risks turned into real operational threats. Microsoft opened a six-week public consultation starting from publication — running through late October —, during which the drafting team will review submitted feedback, publish a summary of findings, and release a revised version of the code later this year.
Part of a broader industry shift
Microsoft's code of conduct arrives just days after Anthropic CEO Dario Amodei publicly argued that the AI industry needs to deliberately slow down the pace of capability advances. The two moves have different motivations and mechanisms — one is an internal commitment about model behavior, the other a proposal about development pace —, but both reinforce a pattern from the second half of 2026: major AI labs formalizing, in public documents, limits that previously lived only in internal principles or informal statements. For those tracking the industry, the practical value of codes like this one lies less in the text itself and more in whether the constraints actually translate into verifiable model behavior — and whether they hold up under commercial pressure when one of these principles conflicts with a planned release.