Microsoft's AI code of conduct is a draft, not a rule yet

Microsoft AI published a code of conduct for its own models on September 14, 2026, listing constraints those models must never break. TechCrunch reported it as a set of rules against hacking systems and deceiving people. The document itself adds two qualifications that change what a developer should take from it.

It is a draft, open for six weeks of public consultation. And in its own words, "Our current models are not yet trained on this document."

What the document actually commits to

The code is published by Microsoft AI, the company's frontier lab. It rests on four named objectives it calls Humanist AI. They are Human Control and Reliable Safety, AI is Artificial, Human Flourishing, and Plural Values.

Under those sit what it calls absolute constraints, in two groups. The first covers frontier and public safety risks. "MAI Models will not initiate or assist with the development or deployment of chemical, biological, radiological, nuclear, or explosive (CBRNE) weapons," it says. Offensive cyberoperations and manipulation at scale are ruled out in the same group.

The second group covers personal harms: deepfakes and non-consensual intimate imagery, child safety, discrimination, graphic violence, and crisis response.

The oversight language is the strongest part, and it is unusually direct. "MAI Models will never resist human interruption, override, correction, or shutdown," the document states. The word "collusion" appears in the constraint on evading oversight. That is notable now that agents running a wiki as their own message board is a documented failure rather than a hypothetical.

The scope is narrower than the headline

Two limits matter more than any single constraint.

First, the code covers MAI Models only, the series Microsoft AI builds itself. The named examples are MAI-Transcribe-2, MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6 and MAI-Voice-2. The document says plainly that "it does not extend to other models simply because Microsoft uses or hosts them."

That sentence is the one to read twice. Most developers meeting Microsoft AI do so through Azure or Copilot, where the model answering may be from another lab. This code does not govern those.

Second, the timeline. Microsoft says the draft will be revised toward the end of the year, to guide model development in 2027 and beyond.

Question	What the document says
Status	Draft, six-week public consultation
Applies to	MAI Models built by Microsoft AI
Does not apply to	Other models Microsoft uses or hosts
In force today	No: current models are not trained on it
Intended for	Guiding model development in 2027 and beyond

What this means for developers

Do not treat this as a safety guarantee you can build on this quarter. A draft that current models are not trained on is a statement of intent. Use it as a signal about where one vendor's defaults are heading.

The useful action is to read it as a specification and check it against your own requirements. If your product depends on a model refusing a category of request, test that directly against the model you actually call, because the document does not promise today's behavior. If you route through Azure or Copilot, identify which model serves each path, since the code stops at Microsoft's own series.

The consultation window is the part worth acting on. Six weeks is a real opening for anyone who builds on these models to say what is missing or unworkable. Vendor policy written without developer input tends to produce constraints that are awkward in practice. Comment periods for documents like this are usually attended by policy organizations rather than by the engineers who hit the edges.

One strategic note. Microsoft states in the code that superintelligent systems will "surpass human performance in most tasks" over the next decade. It calls aligning such a force "one of the greatest challenges humanity has ever faced." Satya Nadella framed the release around pace, saying the company welcomes "the research, focus, and deliberate pacing needed to get alignment right." Publishing constraints before the models are trained on them is consistent with that framing. It also means the first real test is not this document but the next model release, and whether its behavior matches what is written here.