Microsoft's new AI code of conduct says its models must never resist being shut down.

It doesn't say who checks.

The document went out yesterday — thirty-seven pages, open for six weeks of public comment, governing their models from 2027 onward. I read it. It's better than I expected.

The section on honesty is more specific than most companies manage for their own people: don't fabricate sources, don't exaggerate confidence, don't tell users what they want to hear at the expense of what's accurate.

That's a real behavioral standard. You could hand it to an engineer, and they'd know what to build.

Two things aren't in it.

The first is enforcement. Nowhere in thirty-seven pages does it say who verifies any of this, what happens when a model violates it, or how anyone outside Microsoft would know. To their credit, they say so — the document states it "does not substitute for internal or external safety, legal, or governance processes," and that written objectives alone can never ensure alignment.

The second is most of Microsoft's AI. The code binds their own MAI model family. In their words: it "does not extend to other models simply because Microsoft uses or hosts them." Not Azure OpenAI. Not third-party models in Azure AI Foundry. Not Copilot where it runs on something else.

Microsoft is straightforward about both. The headline isn't.

And if you're an operator reading the coverage, you'd think the AI you buy from Microsoft now has a code of conduct. The article is accurate. The conclusion isn't.

Eight weeks ago I finished a series on Asimov, and this is the same shape.

The Three Laws are the oldest AI framework most people can still recite, and they're pure behavioral specification — what the robot must do, nothing about who audits it or what happens when it fails. Asimov knew. He spent forty years writing stories where the Laws are technically satisfied and everything goes wrong anyway, and he invented a troubleshooter whose whole job exists because rules don't enforce themselves.

Eighty-four years later, the newest framework has the same missing half.

In that series I argued responsible AI has to live in three layers at once: what the model's makers build and constrain, the judgment of the person at the keyboard, and the organizational layer in between. This document is the first layer, published at length and in public. That's genuinely useful — it's the layer nobody outside the labs can write.

But it's one of three, and it leaves two questions open.

Who is responsible for enforcing this, and will there be transparency about that enforcement?

And for anyone buying: which framework covers the model you're actually running?

The comment period is open for six weeks. Both seem like reasonable things to ask while it is.