Microsoft says its AI models will never resist being shut down

Microsoft has published a code of conduct for the models it builds itself, and the commitment at the centre of it is blunt. MAI models “will never resist human interruption, correction, or shutdown”. The document went up on Monday alongside a six-week public consultation.

Three further behavioural commitments sit beside it. The models will not widen their own scope, will not take on goals no human has given them, and will not hide their reasoning from the people auditing them.

Above those sit what Microsoft calls Absolute Constraints, things the models should never do at all, covering weapons of mass harm, child safety and harmful manipulation at scale.

The framing is “Humanist AI”, which Microsoft defines as AI that is subordinate, aligned and contained. Its plainest statement of the principle runs: “People matter more than AI. AI should be a tool, not a person, and should never resist being switched off.”

The code covers MAI-Transcribe-2, MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6 and MAI-Voice-2. That list shows how fast the line has moved.

TNW covered the first three in-house models in April, when transcription, voice and image generation were the extent of it, and MAI-Image-2 had just landed third on Arena. A reasoning model and a coding model have arrived since.

One provision complicates the rest. Enterprise partners can configure model behaviour, and Microsoft is explicit that it does not want to impose a single vision of AI on users.

Where configurability stops and the Absolute Constraints begin is the question the document does not answer, and it is the one that determines how much of this is a commitment rather than a default setting.

The consultation runs six weeks, through late October, with a feedback form open to anyone and an invitation to comment on specific passages rather than only the whole. Microsoft says it will publish a summary of what it hears and what it changed, then release a revised version before the end of the year.

On what it will actually accept, the wording is candid:

“We cannot make any promises about what we incorporate, but we can promise to listen.”

The questions Microsoft puts to the public are an admission of what remains unsettled. It wants help embedding values in models, concrete definitions of human flourishing, clearer evaluation criteria, thinking on what happens when multiple agents interact, and views on how to weigh acceleration against safety constraints. Those are the hard parts, and they are being asked rather than answered.

The timing places Microsoft in an argument that went public last week. Dario Amodei said the industry must slow down and Sam Altman matched it within hours, OpenAI’s chief scientist had already said no lab should scale at maximum speed, and Donald Trump dismissed the case for slowing anything.

Microsoft has taken a third position: no pause, no dismissal, a written list of what its models will not do and an invitation to be marked against it.

What the document does not describe is how any of it gets enforced. No external verification is described, no penalty is stated for a model that breaches a constraint, and no account is given of who decides whether one has. Nor is the code attributed to anyone.

Microsoft credits teams from Responsible AI, legal, red teaming, safety, Futures, AI training and sales, which is a revealing list. A code of conduct drafted partly by the sales organisation is a commercial document as much as a safety one, and Microsoft has not pretended otherwise.