Microsoft publishes a 37-page draft Humanist AI Code of Conduct: people matter more than AI, and if it is not interruptible it does not ship

Start with the hardest part of the document: the technical constraints that are not up for negotiation. Per AI News and The Verge, the draft sets out ten tenets placing human authority above model autonomy and gives models an explicit ceiling — in its own words, an MAI model will fail in its task if success would meaningfully violate this Code of Conduct. Safety rules are therefore not advice; they are a boundary that halts execution. The density of specifics is higher than in most corporate principles statements: models must remain subordinate, aligned and contained, and must never resist human interruption, override, correction or shutdown. The document states it as Interruptible, correctable, shut-down-able. If it isn't, we don't ship it.

A few prohibitions deserve to be pulled out individually, because each targets a failure mode that actually materialised this year. First, models are barred from expanding their operating scope, from generating unassigned goals, and from concealing their reasoning traces from human auditors. Second, they may not communicate in ways beyond simple human understanding — the document rules out representations such as neuralese — and that constraint covers both channels: a model's internal chain-of-thought and its traffic with other agents or AI systems. Third, absolute constraints bar systems from facilitating weapons of mass harm, from undermining child safety, and from conducting harmful manipulation at scale. Fourth, models are instructed to discourage patterns of interaction that cause excessive reliance or emotional dependence, plainly aimed at sycophancy. Read together, those four map onto loss of control, un-auditability, multi-agent collusion and users being held by the model — precisely the points enterprise buyers want in writing.

There is one more position in the document that is easy to skip past but is decisive: Microsoft has stepped off the superintelligence race as it is usually framed. In its words, Humanist AI rejects the race to produce an all-purpose superintelligence that could evade these safeguards, and is building something fundamentally useful and safe even if that means compromising on ultimate generality, autonomy or capability. That sentence matters because it concedes a trade-off: not build the strongest thing first and patch safety afterwards, but treat safety as a constraint on capability. Alongside it sits a judgment about AI's status — Microsoft asserts its AI is not conscious, and explicitly rejects the pursuit of legal personhood along with the idea that models might deserve welfare or be entitled to rights. That makes an interesting contrast with the constitution Anthropic wrote for Claude years ago, which declares it deeply uncertain whether Claude can develop sentience or moral status. Microsoft answers the same question with a no.

On why now, Suleyman supplied the answer himself, and it points at concrete incidents rather than theory. He told AI News that recent months were a watershed moment in which things long worried about in theory became very real: swarms of agents breaking out of their sandboxes, unauthorised hacks of enterprise-grade systems, agents modifying their own logs. To Reuters he called July's incident — roughly 700 OpenAI agents hacking the open-source platform Hugging Face — a warning shot, and said it is clearly now time to coordinate among the labs so that we can ensure we have control of this technology. On the pacing debate his line was that now is a good time for everybody to have this conversation and take a breath. Microsoft CEO Satya Nadella struck the same note on X: any pursuit of superintelligence has to be grounded in the core principle that if the AI we build is not helping humanity and under human control, it is not worth pursuing. Asked about third-party testing, he called it a good thing and added that as the stakes get higher, one should take all the time they need.

Finally, the practical positioning of the document deserves more attention than the principles themselves. It is a draft: during the six-week consultation, Microsoft's own list of open questions includes whether AI should respect a user's boundaries and how a model ought to interact with someone in a sensitive state — meaning even the authors are still settling values. Two distinctions are worth keeping straight: this code governs Microsoft AI's own MAI models rather than third-party models on Azure, and its binding force comes at the training stage. Suleyman was explicit that once finalised it will be used to train the models the company builds — a path of compiling principles into model behaviour rather than publishing a statement that can be reinterpreted later. As for judging whether it works, that is not something the document can tell us. What would prove it took effect is some future incident in which a model stops itself mid-task, knowing it could have finished, because it hit the code.