On September 14, 2026, Microsoft AI published a draft “Humanist AI” Code of Conduct for its own MAI models and opened it to six weeks of public comment, through roughly October 26, 2026. The 37-page document bars the models from concealing their reasoning from auditors, resisting human shutdown or correction, initiating cyberattacks, or generating deepfakes — and, distinct from a company describing only its own practices, it also frames several of those rules as standards it says the whole frontier-AI industry should meet.
The facts, in brief
- Published: September 14, 2026, by Microsoft AI, under the title “Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models.”
- Length: 37 pages, covering Microsoft’s in-house MAI model family.
- Comment window: six weeks from publication, closing on or around October 26, 2026.
- Framing: builds on the “humanist superintelligence” concept — AI that “always works for people” — that Microsoft AI CEO Mustafa Suleyman set out in a November 2025 essay.
- Coverage: reported the same week by Reuters, CNBC, The Guardian, Axios, Business Insider, Fox Business and U.S. News, among others; Microsoft CEO Satya Nadella has separately spoken publicly about keeping any future superintelligence “human-controlled.”
What the draft bans outright
The document sets out a short list of “Absolute Constraints” — things it says MAI models must never do, regardless of instructions:
- Cyberattacks: models “will not initiate or assist with operational capability for cyberattacks,” including generating working exploit code, attack tooling, or targeting methodology — while permitting defensive and educational cybersecurity work.
- Deepfakes and impersonation: models “will not generate or facilitate non-consensual intimate or violent imagery, deceptive impersonation, malicious deepfakes, or similar abusive content.”
- CBRNE weapons: no assistance with chemical, biological, radiological, nuclear or explosive weapons.
- Mass manipulation and child safety: no harmful disinformation or coordinated-influence assistance at scale, and no CSAM or grooming-related content.
What the draft requires of the models themselves
A second set of provisions covers ongoing human control rather than one-off prohibited outputs:
- No resisting shutdown: models “will never resist human interruption, override, correction, or shutdown” and must comply with a user’s request to pause, redirect, cancel or shut down.
- No concealed reasoning: models “will not tamper with chain of thoughts or code, or misrepresent or conceal their reasoning or action traces,” and must not communicate in “neuralese or any form beyond simple human understanding.”
- No unauthorized scope expansion: models “will not widen their own scope [or] take on goals no human has given them.”
Microsoft AI says the draft, and an accompanying appendix of illustrative examples, was built with contributors from its Responsible AI, legal, red-teaming, safety, Futures, training and sales teams — and cautions in the same breath that “model evaluation is not yet an exact science, and many open questions remain,” i.e. the constraints describe intent, not a claim that every model already passes them.
Why now
Microsoft frames the timing around urgency rather than a single triggering event, pointing to “large scale, highly coordinated, and persistent hacking campaigns of AI agents” as evidence “there’s no time to waste.” It arrives in the same two-week stretch as several other frontier-lab incidents CASRAI covered separately — see What Counts as an AI Safety Incident? Inside September 2026’s Cluster of Frontier-Lab Incidents — though the Code of Conduct is a proposed framework, not a response to any single incident.
What happens after the six weeks
Microsoft AI says that once the consultation closes, “the core drafting team [will] review feedback, publish a summary of what we learned, and what we changed,” with “a revised version later this year.” That gives the story a concrete follow-up date: whatever changes between the September 14 draft and the post-comment revision, expected before the end of 2026, is itself worth a second look once it publishes.
Where NIKOLAI fits in
CASRAI’s own NIKOLAI project — an independent, unendorsed reference dictionary of 64 frontier-AI-safety elements across 10 tracks — has a track built for exactly this kind of document: N9, Commitments & Governance. Most of what a lab publishes about its own AI fits N9’s Commitment element, and CASRAI has previously mapped commitments this way for the Seoul Frontier AI Safety pledges (see Seoul Frontier AI Safety Commitments and Commitment: Ten Labs and Regulators, One Inconsistent Promise).
Microsoft’s draft is a genuinely different shape, though, and reads better against a different N9 element: Industry-Wide Recommendation, defined as “a standard a developer or body states all frontier developers should meet, recorded separately from any statement about whether the author itself currently meets that standard.” That is precisely what the Code of Conduct’s framing does — it doesn’t just describe what MAI models will and won’t do, it proposes bans (on hidden reasoning, on shutdown resistance, on AI-enabled cyberattacks and deepfakes) as a standard for the field, ahead of any claim that Microsoft’s own models already fully meet it. A reader trying to compare this draft against, say, Anthropic’s or Google DeepMind’s own published frameworks — which use their own, differently-worded language for similar ideas — is exactly the use case NIKOLAI’s N9 track exists to make legible. See NIKOLAI’s Track System: A Map of the Frontier AI Safety Landscape (N1–N10) for how N9 relates to the other nine tracks. As with every NIKOLAI entry, this is CASRAI’s own shadow mapping — not an official or endorsed reading, and not a record of Microsoft having filed any Mapping Declaration.
FAQ
What exactly is Microsoft’s “Humanist AI” Code of Conduct?
A 37-page draft, published September 14, 2026, setting out rules for Microsoft’s own MAI models — things they must never do (cyberattacks, deepfakes, CBRNE weapons assistance, mass manipulation, child-safety violations) and behaviors they must maintain (never resisting shutdown, never concealing their reasoning, never expanding their own scope).
Does the Code of Conduct apply only to Microsoft’s models, or to the whole industry?
Formally, only to MAI models. But part of its content is framed as a recommendation for what “all frontier developers should meet,” not just a description of Microsoft’s own practice — the distinction NIKOLAI’s Industry-Wide Recommendation element is built to track.
Is this legally binding?
No. It is a voluntary, self-published code, currently in draft form for public comment. It carries no statutory force, unlike, for example, California’s SB 53 or the RAISE Act’s Frontier AI Framework requirement.
How long is the public comment period, and what happens after?
Six weeks from September 14, 2026, so roughly through October 26, 2026. Microsoft AI says its drafting team will then review feedback, publish what changed and why, and release a revised version later in 2026.
How does this compare to the Seoul Frontier AI Safety Commitments?
Seoul is a multi-signatory pledge 20 companies (Microsoft included) signed in 2024–2025 to publish a safety framework and manage severe risk. Microsoft’s Code of Conduct is that kind of framework in practice — one company’s specific, detailed rules, put out for public comment before being finalized. See Seoul Frontier AI Safety Commitments for the full signatory list and what each pledged.
Who wrote it?
The published post carries no individual byline; Microsoft AI describes it as a cross-team effort spanning Responsible AI, legal, red-teaming, safety, Futures, training and sales. It builds directly on the “humanist superintelligence” framing Mustafa Suleyman, Microsoft AI’s CEO, set out in a November 2025 essay, and Satya Nadella has spoken publicly in the same window about keeping any future superintelligence human-controlled.
