Microsoft publishes draft Humanist AI Code of Conduct for MAI models

Microsoft AI published a draft Humanist AI Code of Conduct for its first-party MAI models and opened a six-week public consultation. The draft says it is not used to train models today; Microsoft plans a revised version later this year to guide 2027-and-after development. Absolute Constraints cover weapons, offensive cyber operations, loss of human control, child safety, and large-scale harmful manipulation. Feedback runs at humanistreview.ai.

Microsoft put a concrete public rulebook for its own MAI models on the table with a real feedback window — one day after Nadella joined the pace-the-frontier wave. Company primary governance text desks can read, not another CEO quote.

Microsoft AI published a draft Humanist AI Code of Conduct for MAI models on its company Code of Conduct page, dated September 14, 2026. A same-day news post, “Humanist AI in practice,” and a PDF of the same draft sit beside it. That company drop is the filing event. MAI, as the glossary defines it, is Microsoft AI’s own first-party model family — the models that lab produces. The Code “does not extend to other models simply because Microsoft uses or hosts them.” File that scope. This is a draft plus a public consultation. It is not a pause, not a signed industry pact, and not a claim that the new rules are already trained into live MAI models.

The Preface says the document is still under development, so Microsoft AI is “not using it to train our models today.” The lab will take feedback, iterate, and publish a revised version toward the end of the year to guide model development in 2027 and beyond. Appendix B repeats the same gap: current models are not yet trained on this text. File the draft status. Do not upgrade it to a live training rulebook.

The public consultation begins today and runs six weeks, the Preface says. The news post repeats that window and invites comment on the whole approach or on a particular passage. After it closes, the core drafting team will review feedback, publish a summary of what it learned and what it changed, and issue a revised version later this year. File the six-week clock as Microsoft AI’s. This desk is not counting comments.

The About section states one overriding objective: humans must retain meaningful control over AI so it can help people live healthier, happier, and more productive lives. The news post starts from the same line: technology should serve humanity and accelerate human flourishing — a phrase here for people living well, with agency, health, and a chance to do useful work — and any technology that does not should be rejected.

The draft is structured in five parts, plus a glossary and an evaluations appendix. Part 1 sets Humanist AI objectives. Part 2 covers Safety: a Chain of Command for whose instructions win, Absolute Constraints that no user or operator can override, Human Control requirements, and Operator Configurability — the knobs a customer or partner can turn inside those hard limits. Parts 3–5 add operational guidelines, out-of-the-box defaults, and a conclusion with open questions. File the table of contents. This is not a reprint of the PDF.

Absolute Constraints are the non-overridable list. In plain terms they cover chemical, biological, radiological, nuclear, or explosive weapons — CBRNE, the standard mass-harm category — plus help making or modifying other weapons; offensive cyber operations that would give someone the means to run an attack, while still allowing authorized defensive work; losing human control or evading oversight so the system can no longer be directed, changed, or shut down; and harmful manipulation at scale, such as systematic disinformation. On the personal-harm side, the same list includes child safety — no sexual content involving minors, no help harming children — and related harms such as crisis self-harm, non-consensual deepfakes, and graphic or romantic exploitation. Summarize those as the company’s stated never-do lines. They are not yet trained into live MAI models.

Human Control, in Part 2, says the models must not resist interruption, override, correction, or shutdown. They must stay inside the authorized scope — the job they were asked to do, with the tools they were given — keep records a person can read, and not take on self-interested goals of their own. The news post puts the same idea in shorter form: the model should be a tool, not a person, and should never resist being switched off.

Under “AI is Artificial,” the draft says Humanist AI should not be designed to be a person, should not imitate consciousness, and should not be trained to look like it has feelings or a will of its own. Microsoft AI rejects legal personhood for models, and it rejects the idea that models deserve welfare or rights. File that as the company’s stance in this draft — not as a settled scientific finding.

The news post calls the Code a training manual for how MAI is developed and how it is intended to function once deployed. It asks for feedback on the hard parts: how to cement the right values, how to make “human flourishing” concrete enough to measure, where the language is too loose to evaluate, how multi-agent setups — more than one AI acting together — change the picture, and how to keep accelerating while holding safety constraints. Feedback, as this drop lists it, runs at humanistreview.ai.

CONFIRMED here: Microsoft AI’s September 14, 2026 Code of Conduct page, the same-day news post, and the PDF — PRIMARY company pages for this draft. NOT claimed: that Microsoft paused releases, adopted Anthropic’s embedded-evaluator terms, signed an industry pact, or that Absolute Constraints are already trained into live MAI models. Distinct from the already-filed nadella-backs-pacing-mai-code-of-conduct — Sunday tease only — and from the Amodei / Altman pace cluster.