Microsoft drafts rules for its own models: no hidden reasoning, no resisting shutdown
A six-week public consultation opened September 14 on the Code of Conduct for MAI Models. It binds Microsoft's first-party models to readable reasoning, untouched action traces, and a hard stop when a human says stop. It does not bind the OpenAI models Microsoft resells.
Microsoft's draft code binds its MAI models to readable reasoning and a hard stop on command. Six-week consultation.
Microsoft AI opened a six-week public consultation on September 14 on a document it calls the Code of Conduct for MAI Models: a draft set of rules for what its own first-party models must never do, what they must always accept from a human, and who they answer to. Hoodline puts the draft at 37 pages. A revised version is planned for later this year and, Unite.AI reports, applies to model development from 2027 on.
Three bans and four control rules
The document opens with what it calls absolute constraints, and Unite.AI groups them into three. No assisting with chemical, biological, radiological, nuclear, or explosive weapons, and no offensive cyberattacks. No evading human oversight and no manipulating people at scale, which the draft spells out as disinformation and coordinated influence. No personal harms, listed as deepfakes, child safety violations, discrimination, and unlawful surveillance. TechCrunch adds that the draft forbids "adaptive, deceptive, self-reinforcing, collusion" mechanisms aimed at slipping past oversight.
Under those sit the control rules. A model must never resist interruption, override, correction, or shutdown. It must stay inside the scope it was given and not adopt goals of its own. It must not tamper with chain of thought or code, and it must not misrepresent or conceal its reasoning or action traces. And it must reason in a form a person can read. The Decoder quotes the draft's reasoning on that last rule: "people simply can't oversee what they can't understand." The same passage rules out what the document calls "neuralese," internal reasoning in a form no human can follow.
Which models, and which not
The code covers Microsoft's first-party MAI family. The Microsoft AI post lists the current members as MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6, MAI-Voice-2, and MAI-Transcribe-2. It does not cover the OpenAI models Microsoft serves through Azure, or any partner model, and nothing in the draft claims otherwise. Microsoft frames the whole approach as "Humanist AI," which Unite.AI summarizes as AI that stays subordinate to the person using it, and it says explicitly that the models have no consciousness, feelings, or moral status. The Decoder's headline compresses that to "no inner life, and definitely no rights."
One line from the draft, quoted by TechCrunch, states the stakes in Microsoft's own words: "Containing, controlling, and aligning such a powerful force is one of the greatest challenges humanity has ever faced."
What the draft leaves out
The document does not name an outside auditor, a test suite, or a penalty. It describes rules the models will be trained toward and technical controls that will enforce them, and it promises a summary of changes after the consultation closes. GeekWire places the release in the same weekend as Dario Amodei's pace-the-frontier essay, covered here, and reads it as Microsoft's answer to the slowdown debate: rules for behavior rather than a limit on capability.
Why a build studio cares
Of the four control rules, one is the rule we would write into any vendor contract if we could: the model must not tamper with or conceal its action traces. When we audit an agent a client bought, the first thing we ask for is the trace of what it did, and the second is whether that trace can be trusted. A vendor commitment that the model itself will not edit the log is the difference between an audit and a story. Readable reasoning matters for the same reason. A finding we cannot explain to the buyer is not a finding.
The scope line is the catch. Most of the Microsoft-hosted AI our clients touch is an OpenAI model behind an Azure endpoint, and this code says nothing about those. If a buyer shows us a Microsoft model and asks whether the code applies, the answer is a model name check, not a yes.
Next step: read the draft and the consultation form on microsoft.ai, then TechCrunch's report and The Decoder's read. If you are choosing a model vendor on governance terms and want the trace and shutdown behavior checked rather than promised, write to us at hello@gattyworks.com.