Confirmed in a primary source 14 Sept 2026 Safety

Microsoft published the rulebook it promised for its own AI models, and it binds nothing yet.

On Sunday Satya Nadella said Microsoft welcomed a slower pace for AI and would publish, the next day, a code of conduct for the models Microsoft builds in house. On Monday the document appeared: 38 pages setting out that a Microsoft model must never resist being interrupted, corrected or shut down, must not hide its reasoning or talk to other models in formats people cannot read, must not claim legal personhood, and must refuse weapons help, child sexual content and manipulation.

The caveat is written into the document itself. It applies only to Microsoft's own MAI models and explicitly not to other models Microsoft uses or hosts, which is where most of the AI inside Microsoft products comes from. It says it is not being used to train anything today. A revised version is due at the end of the year to guide development in 2027 and beyond. Comments close on October 25, and Microsoft says it can make no promises about what it takes in.

Until now

The weekend's pacing endorsements from four chief executives were words, and Microsoft's was the only one with a delivery date attached to it.

Now

One of the four delivered a document on the day promised. Read it as a scope statement rather than a safety commitment: it covers the models Microsoft builds, not the ones it ships, and it trains nothing until 2027.

Nadella's Sunday post was the fourth chief executive endorsement of pacing in 48 hours and the only one that named a date. The date was Monday, and on Monday the document went up. That is worth saying plainly, because most safety announcements in this industry do not survive contact with a calendar. What went up is 38 pages of behavioural rules with some genuinely hard edges in them. A model may not resist interruption, correction or shutdown. It may not hide what it is thinking, or speak to another model in a format a person cannot read. It may not take on goals nobody gave it. It may not claim to be a legal person. If succeeding at a task would meaningfully break the code, the document says, the model fails the task instead.

Then you read the scope paragraph. The code governs MAI Models, the ones Microsoft AI builds itself, and the text goes out of its way to say it does not extend to other models simply because Microsoft uses or hosts them. Most of the AI a Microsoft customer touches is an OpenAI model running inside Copilot or Azure, and it is not covered. The document also says it is not being used to train Microsoft's models today; a revised version arrives at the end of the year, for 2027 and beyond. Microsoft's Enterprise AI Services Code can cost a customer access to a product for breaking it. This one names no consequence for anybody. One of the four delivered a document on the day promised; read it as a scope statement rather than a safety commitment, because it covers the models Microsoft builds, not the ones it ships, and it trains nothing until 2027.

How it got here

Sep 12
Dario Amodei publishes his pacing essay and commits Anthropic to embedded outside evaluators. Sam Altman, Elon Musk and Demis Hassabis agree within hours. Carded in Edition 020.
Sep 13
Satya Nadella endorses deliberate pacing, backs embedded outside evaluators and says Microsoft will publish its MAI code of conduct the next day for public consultation. Carded as Reported in Edition 021.
Sep 14
The draft Humanist AI Code of Conduct is published at microsoft.ai with a six-week consultation closing October 25. The Edition 021 card's verification event fires on the day it was set.
Sep 14
The same day, the US President rejects new AI guardrails and attacks Amodei by name, and China's state newspaper calls the slowdown call a Cold War tactic.
Who has to react
OpenAI, whose models sit inside Microsoft products and are explicitly outside this code, and which said on September 12 it would accept embedded evaluators; Anthropic, whose unilateral commitment now has a rival artefact to be measured against; Microsoft's enterprise customers, who have a written behavioural standard for one class of model and none for the other; the EU AI Office and the California regulators, for whom a published code is a document they can hold a company to later; and every lab now expected to produce something similar, because a 38-page draft on a promised date is the new floor.

The evidence

ObservedMicrosoft AI published the draft Humanist AI Code of Conduct at microsoft.ai/code-of-conduct on September 14, 2026, with an accompanying post at microsoft.ai/news/mai-code-of-conduct opening a six-week public consultation. We read both.

The whole observation, as recorded

The preface is dated September 14, 2026. The Register counted 38 pages. Scope, in the document's own words: the code sets the intended behaviour of MAI Models, defined as the series of models produced by Microsoft AI used across a number of Microsoft products and services, and it does not extend to other models simply because Microsoft uses or hosts them. Structure is four parts. Part 1, Humanist AI Objectives: Human Control and Reliable Safety, AI is Artificial, Human Flourishing, Plural Values. Part 2, Safety: Chain of Command, Absolute Constraints, Human Control Requirements, Operator Configurability. Part 3, Operational Guidelines, eight headings from resolving conflicts and ambiguity through supporting the public interest. Part 4, Operational Defaults, five headings from helpfulness to tool use. Specific provisions: a model must never resist human interruption, override, correction or shutdown; it must not widen its own scope, take on goals no human has given it, or hide its reasoning; it must not use formats incomprehensible to humans, in internal processing or in model-to-model communication; it must not claim legal personhood; and the text states that an MAI Model will fail in its task if success would meaningfully violate the code. The document says the code is not yet used to train Microsoft's models, that a revised version will be published toward the end of the year to guide model development in 2027 and beyond, and that it does not substitute for internal or external safety, legal or governance processes. Feedback runs through a Microsoft Forms link; the consultation closes October 25, 2026, and the post says Microsoft cannot make any promises about what it incorporates. Mustafa Suleyman described recent months as a watershed moment in which theoretical risks became operational. No outside evaluator is named for ongoing assessment; the consultation lists academics, business partners and panels of community members. Microsoft has a separate Enterprise AI Services Code that can cost a customer product access for violations; this one carries no stated consequence.

Established

Microsoft AI published a draft Humanist AI Code of Conduct on September 14, 2026, the day after Satya Nadella said it would appear, and opened a six-week consultation closing October 25, 2026.
The code states that MAI models must never resist human interruption, override, correction or shutdown, must not hide reasoning or use formats incomprehensible to humans, and must not claim legal personhood.
The code applies only to models produced by Microsoft AI and explicitly does not extend to other models Microsoft uses or hosts.
The document states it is not yet used to train Microsoft's models and that a revised version is planned toward the end of 2026 to guide development in 2027 and beyond.

Not established

That any Microsoft model behaves this way today. The document says it is not being used to train models now, and publishes no evaluation showing current compliance.
That the code will be enforced. No penalty, auditor, escalation path or reporting obligation is attached, and Microsoft's separate enterprise code, which does carry consequences, is a different document.
That Microsoft has accepted embedded outside evaluators. Nadella endorsed the idea on September 13; this document names no external evaluator and creates no access.
That this covers the AI inside Microsoft products. The OpenAI models running in Copilot and Azure are outside the stated scope.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityModerate

PersistenceYear

Cost is medium. Writing and publishing 38 pages of behavioural rules is real legal and drafting work and it creates a public text that critics and regulators can quote back, which is a genuine cost. It is not high, because no money moved, no product changed and nothing was given up. Commitment is medium and deliberately bounded: the document commits Microsoft to a consultation and a revised version at the end of the year, and it commits to nothing enforceable in the meantime, saying in its own text that it is not being used to train models today and cannot promise what it incorporates. Reversibility is moderate rather than easy, because a published draft is hard to retract quietly, but every specific provision can be softened in the revision due in about three months. Persistence is a year: the revised code is meant to guide development in 2027 and beyond, so the commitment has a horizon rather than being structural.

What to watch

strengthens A second frontier lab publishes a comparable model-behaviour code or constitution for public consultation, or Microsoft extends this code to cover models it hosts rather than builds, by December 31, 2026.
weakens The consultation closes on October 25, 2026 and no revised code, no summary of responses and no named outside evaluator is published by March 31, 2027.

The words

MAI models: The models Microsoft AI builds in house, as opposed to the OpenAI models Microsoft licenses and runs inside Copilot and Azure. The code covers only the first kind.
Code of conduct: Here, a written statement of how a company intends its models to behave. It is not a law, a contract or a licence condition, and this one names no penalty for breaking it.
Public consultation: A period in which anyone can send comments on a draft. Microsoft's runs six weeks and closes October 25; it has said it will publish findings and a revised draft.
Chain of command: The document's term for whose instructions a model follows first when they conflict: Microsoft's rules, then the business deploying the model, then the user.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.