Microsoft opens six-week public comment on draft AI code of conduct
Decrypt
Microsoft just opened a public comment on its AI code of conduct. It bans weapons help, cyberattacks, and deepfakes — and says humans stay in charge.
Based on reporting by Decrypt — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Microsoft AI has published a draft code of conduct for its own models and is asking the public to weigh in for six weeks, until late October. The document, announced by Mustafa Suleyman on X, lays out what the company calls hard limits and broader goals for how MAI models should behave.
The hard limits are the eye-catching part. Microsoft says its models can’t be configured to help with CBRNE weapons — chemical, biological, radiological, nuclear, or explosive — and separately bars cyberattacks and nonconsensual deepfakes. The draft also says MAI models must never resist human interruption, correction, or shutdown. Human intent, in Microsoft’s telling, always comes first.
There’s another notable line in the sand: Microsoft rejects “model welfare” as a premise. Its systems are not supposed to simulate feelings, intrinsic motivation, or consciousness. The current rules apply to five deployed systems, including MAI-Thinking-1 and MAI-Code-1.1-Flash.
But the draft is not just a list of prohibitions. It also sets out three softer objectives — Human Flourishing, Plural Values, and Human Control — for situations where the rules don’t settle the answer. Plural Values is framed as a rejection of neutrality toward harm, with decisions meant to be grounded in dignity, safety, autonomy, and rights. What that means in practice is still being worked through with training teams and, now, outside commenters.
Microsoft is also making clear this is not finished policy. The company says current models are not being trained on the draft, and the code will not shape MAI development until a revised version arrives toward the end of the year, in time to guide 2027 releases. The consultation may also fill in gaps that are still missing, including external verification and a named enforcement owner.
My take — AI-written commentary, not fact-checked reporting
This is the classic AI-lab move: write a moral manifesto after the model is already out in the wild, then ask the public to help clean up the mess. The real tell is the part about “human control” — companies love saying that right up until a product gets too useful to slow down. The industry has turned restraint into branding, which is a very modern way to sell the brakes while keeping the engine running.
Read more about this at: Decrypt