Reported 13 Sept 2026 Safety

Microsoft's chief executive backed outside evaluators for AI and said the company would publish rules for its own models the next day.

On Sunday Satya Nadella posted that Microsoft welcomes the deliberate pacing that alignment work needs, endorsed embedded outside evaluators and mechanisms to substantiate what labs claim about alignment, and said Microsoft would publish the next day, for public consultation, the code of conduct that underlies its own first-party MAI models. That puts the largest enterprise seller of AI behind the position Anthropic took on Saturday and OpenAI matched within hours, and unlike either of them it attaches a dated, checkable deliverable.

The caveat is that the deliverable had not appeared when this was filed: the newest items on Microsoft's AI newsroom and its policy blog are September 4 and September 10, the American working day had not begun, and an endorsement plus a promise is still an endorsement plus a promise.

Until now

The pacing turn was a frontier-lab conversation: two model builders promising outside inspectors, with nothing written down and no deadline on anything.

Now

Treat the pacing turn as having reached the distribution layer, where the commitment on offer is a published document opened to comment rather than an inspector given a desk, and judge each company by which of the two it actually produces.

The sequence now runs to four. Dario Amodei published on Saturday and committed Anthropic to letting outside evaluators work inside the company with the access its own risk staff have. Sam Altman said OpenAI would do the same. Elon Musk said Amodei was right, and Demis Hassabis agreed the same evening. On Sunday Satya Nadella added the largest of the distributors: Microsoft welcomes deliberate pacing, supports embedded evaluators, supports mechanisms that let alignment claims be substantiated, and will publish for public consultation the code of conduct behind its own MAI models. It is a different kind of company making a different kind of promise. Microsoft is not principally a frontier lab; it is the channel through which most enterprises buy this technology, and what it governs is what its models will and will not do for a customer.

The promise is also the only one in the sequence with a clock on it. Amodei named no evaluator and no start date, Altman matched a commitment with no terms, and Nadella said tomorrow. Tomorrow is today, and as of the moment this was filed the document does not exist: nothing dated September 13 or 14 appears on Microsoft's AI newsroom or its policy blog, whose newest posts are September 4 and September 10, and the American working day has not begun. That is not a failure yet, it is simply the state of the evidence, and it is why this card is Reported and why the check is dated to the day. Note also what is missing from the post. Microsoft backed embedded evaluators without committing to host any, which leaves the strongest form of the pledge with the two labs that made it. Treat the pacing turn as having reached the distribution layer, where the commitment on offer is a published document opened to comment rather than an inspector given a desk, and judge each company by which of the two it actually produces.

How it got here

Sep 12
Dario Amodei publishes "We Must Pace the Frontier" and commits Anthropic to embedded third-party evaluators with staff-level access and the right to publish uncleared findings. Sam Altman says OpenAI will do the same; Elon Musk and Demis Hassabis agree. Carded in Edition 020.
Sep 12
Hugging Face announces an Open Alignment Initiative under Thomas Wolf and asks to be one of the evaluators. Recorded in Edition 020's Undercurrents.
Sep 13
Satya Nadella posts Microsoft's endorsement of deliberate pacing and embedded evaluators and says the MAI code of conduct will be published the next day for public consultation.
Sep 14
As of 06:00 UTC no code of conduct has appeared on any Microsoft property. Asian markets open and sell AI exposure.
Who has to react
Microsoft's enterprise customers, who will read a model code of conduct as a procurement document and comment accordingly; OpenAI, whose models Microsoft sells and which this document does not cover; Anthropic, whose stronger commitment is now the one Microsoft declined to match; the evaluation organisations that would have to staff any of this, METR and Apollo Research among them, none of whom have been named by anyone; European regulators drafting AI Act codes of practice, who now have a voluntary corporate text to compare against their own; and Senate negotiators reported to be drafting a duty of care for frontier developers.

The evidence

Nadella's post is on X and dated September 13; the endorsement and the publication promise are established. The document itself is the open question. As of 06:00 UTC on September 14 nothing dated September 13 or 14 appears at microsoft.ai/news or on Microsoft's On the Issues blog, whose newest items are September 4 and September 10. Redmond had not opened for business when this was filed. The verification event is the code of conduct appearing on a Microsoft property with a public consultation process attached, on September 14, 2026.

ObservedSatya Nadella posted on X on Sunday September 13, 2026. The quoted commitments: Microsoft "welcomes the deliberate pacing needed" for alignment work; support for embedded evaluators and for mechanisms to substantiate alignment commitments; "broad access and choice at every layer of the AI stack"; enterprise control of learning loops and of model weights; a frontier ecosystem supporting both closed and open-source models.

The whole observation, as recorded

The line carried by every outlet: "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's not worth pursuing." The specific undertaking: Microsoft would publish "tomorrow for public consultation" the code of conduct that underlies its own first-party MAI models. Those models, as listed on Microsoft's AI newsroom, include MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6, MAI-Transcribe-2 and MAI-Cyber-1-Flash, distributed through Foundry and third-party routers. What Microsoft did not say: it did not commit to embedded evaluators inside Microsoft with the access Anthropic described, it named no evaluator, it set no enforcement mechanism, no consultation closing date and no safety-standard timeline. Verification as of 06:00 UTC on September 14: microsoft.ai/news carries nothing newer than September 4, 2026, and blogs.microsoft.com/on-the-issues nothing newer than September 10, 2026. This is the fourth chief executive endorsement in 48 hours, after Sam Altman, Elon Musk and Demis Hassabis on September 12, and it is the first to carry a date.

Established

Satya Nadella publicly endorsed deliberate pacing and embedded outside evaluators on September 13, 2026, and stated that Microsoft would publish a code of conduct for its first-party MAI models the following day for public consultation.
This is the fourth public endorsement of the pacing position by a major AI chief executive within 48 hours.
The endorsement covers Microsoft's own MAI models, not the OpenAI models Microsoft distributes.
As of 06:00 UTC on September 14, 2026, no such document appears on microsoft.ai/news or blogs.microsoft.com/on-the-issues, whose newest items are September 4 and September 10.

Not established

That Microsoft will host embedded evaluators. Nadella supported the idea and committed Microsoft to nothing of the kind.
That the document will say anything binding. No draft exists publicly, and a consultation has no obligation to produce a rule.
That the industry has agreed on anything. Four endorsements in 48 hours with no named evaluator, no terms and no start date is agreement about a direction, not a standard.
That anyone has slowed down. No model release has been delayed, no training run paused and no product withdrawn by any of the four companies.

Why it weighs this much

Signal strength

CostLow

CommitmentMedium

ReversibilityEasy

PersistenceMonths

Cost is low and we are not going to pretend otherwise: a post on X and a promise to publish a document costs nothing, and Microsoft has committed no money, no staff and no access to anyone. Commitment is medium rather than low for one reason only, that a named next-day publication date is a thing that can be checked and missed in public, which is more than any of the three endorsements before it carried. Reversibility is easy: a consultation draft can be withdrawn, softened or quietly never finalised. Persistence is months, because a published code of conduct with comments attached becomes a document people cite back at you, which outlasts a post, but it binds nobody and no regulator has adopted it.

What to watch

strengthens By December 31, 2026, at least two of Anthropic, OpenAI, Microsoft and Google DeepMind name a specific external evaluation organisation with published terms of access, or a joint industry standard is published with named signatories.
weakens Microsoft does not publish the MAI code of conduct by September 30, 2026, or publishes it with no stated consultation process or closing date.

The words

Embedded evaluator: An outside safety team given working access inside a lab. Anthropic's version, announced September 12, means desks, badges, company laptops and the right to publish findings the company has not approved.
Code of conduct: Here, a published document setting out how a company's models are meant to behave and what they refuse to do. It binds the company only as far as the company chooses to honour it.
Public consultation: Publishing a draft and inviting anyone to comment before it is finalised. Common for regulators, rare for a company's own model rules, and it creates a dated record of who objected.
MAI models: Microsoft's own family of models, built in-house rather than licensed from OpenAI. The code of conduct is said to cover these, not the OpenAI models Microsoft also sells.

Your call

If the code of conduct is published today as promised, which explanation do you believe?

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