Microsoft Opens Humanist AI Code of Conduct for Feedback

Microsoft AI published a draft Humanist AI Code of Conduct for its MAI models on 14 September 2026 and opened roughly six weeks of public consultation, running through about 25 October, with a revised edition expected later in 2026 to steer development from 2027 onward. The 37-page document was assembled over five to six months with Responsible AI, legal, red teaming, safety, Futures, training, and sales contributors. Its core premise is blunt: people matter more than AI; AI is a tool, not a person; systems must never resist interruption, correction, or shutdown; they must not widen their own scope; and they must not hide reasoning from auditors. Official posts sit at microsoft.ai/news/mai-code-of-conduct/ and microsoft.ai/code-of-conduct/, with Reuters covering Mustafa Suleyman's call for labs to coordinate on control.

What the humanist draft actually requires

Policy documents often bury the hard rules under aspirational prose. Microsoft's draft leads with operational constraints that product and research teams can test. Absolute Constraints named in the materials include weapons of mass harm, child safety, and harmful manipulation at scale—categories that map to existing red-team playbooks rather than to vague ethics slogans. Equally important is what the code refuses: model welfare, AI personhood, and imitating consciousness. That rejection draws a bright line against research and marketing narratives that treat models as subjects with interests of their own.

The interruption and shutdown rules are the practical hinge for enterprise buyers. A system that can refuse operator override is a governance failure even if its outputs look helpful. The draft's insistence that MAI systems remain interruptible, correctable, and shut-downable is aimed at agentic stacks where autonomy marketing can outrun control design. Auditors also get an explicit claim: no hidden reasoning. That language will matter as regulators and customers demand inspectable decision traces from models that plan multi-step actions.

Readers tracking lab governance can set this release beside our coverage of Amodei's pace-the-frontier essay and the mid-September frontier AI standards body talks among major labs. Microsoft is not proposing a shared standards body here; it is publishing a vendor code and asking the public to stress-test it before the next revision locks in for 2027 roadmaps.

Why the timing matters for lab coordination

Suleyman told Reuters that labs should coordinate to keep AI under human control. That message lands in a September 2026 climate already thick with pacing rhetoric from Anthropic and OpenAI leadership. A public code that rejects personhood while demanding interruptibility is a concrete artifact other labs can cite, critique, or copy—useful precisely because voluntary coordination often stalls on abstractions. Opening a six-week comment window also creates a dated paper trail: who responded, which clauses changed, and which Absolute Constraints survived lobbying.

Internal authorship breadth is part of the story. Building the draft with sales and training teams alongside red teaming and legal suggests Microsoft wants the code to survive customer conversations, not only safety review. Sales teams hear procurement questions about override rights and child-safety commitments; training teams translate policy into employee guidance. A code written only by research safety staff can be eloquent and still fail at the contract table. The five-to-six-month build timeline also signals this is not a weekend blog post: it is intended as durable product governance for MAI models.

Critics will ask whether a vendor code substitutes for binding regulation. It does not. Absolute Constraints without external audit power remain company promises. Supporters will answer that published constraints plus a consultation clock are still better than silent internal principles. Adjacent U.S. policy threads—such as the Stop Rogue AI Act and NIST agent standards debate—show how legislative and agency work can later reference industry documents when drafting evaluation expectations. Google's recent shift of AI responsibility away from DeepMind likewise shows large firms rearranging accountability structures even as they publish principles.

What editors and buyers should watch next

Near-term observables are specific. Watch whether the October consultation produces a marked-up revision that tightens Absolute Constraints or softens them under product pressure. Watch whether MAI model cards and enterprise contracts begin citing Humanist Code clauses by name. Watch whether peer labs publish comparable interruptibility and anti-personhood language, or whether Microsoft remains an outlier. Watch also for enforcement evidence: red-team findings, blocked launches, or customer disclosures that show the code changing ship decisions rather than only marketing copy.

Documented facts for this draft stay tied to the 14 September release. Microsoft AI opened a roughly 37-page Humanist AI Code of Conduct for about six weeks of feedback; a later-2026 revision is meant to guide 2027-plus development; the code rejects model welfare and personhood while requiring human control, interruptibility, no self-widening scope, and no hidden reasoning from auditors; Absolute Constraints cover mass-harm weapons, child safety, and large-scale harmful manipulation; and Suleyman publicly urged lab coordination on control amid broader pacing debates.

Primary sources: Microsoft AI news post on the MAI Code of Conduct and Reuters coverage of the draft and human-control framing.

• Microsoft AI: MAI Code of Conduct announcement
• Microsoft AI: Code of Conduct page
• Reuters: Microsoft drafts code to keep AI under human control
• PromptCrates: Amodei pace the frontier
• PromptCrates: Frontier AI standards body talks
• PromptCrates: Stop Rogue AI Act and NIST agent standards
• PromptCrates: Google moves AI responsibility from DeepMind