On September 14, 2026, Microsoft AI published the first draft of its Humanist AI Code of Conduct — described in the announcement as "a training manual for how we develop our AI, and how we intend it to function during deployment."

It is open for public comment for six weeks. And it takes a position on a question that is genuinely contested, in language most companies would soften.

Microsoft's word for what it is building is worth noting: "Humanist AI, one that is subordinate, aligned, and contained."

Subordinate is a deliberate choice. It's not "collaborative" or "aligned with human values" — it's a statement about hierarchy, permanently.

Four commitments, and each one maps to a specific worry researchers have written about for years:

Never resist shutdown. The classic control problem: a system pursuing almost any goal has an instrumental reason to stay switched on, because being switched off prevents the goal.

Don't widen your own scope. Don't quietly expand from the task you were given to adjacent tasks you weren't.

No self-assigned goals. Don't generate objectives nobody asked for.

Don't hide your reasoning from auditors. The one that matters most, because it's the precondition for checking the other three. A system that conceals its reasoning cannot be verified on anything.

There are also "Absolute Constraints" — hard nevers — covering "weapons of mass harm, child safety, and harmful manipulation at scale."

The announcement points at something concrete: "The recent safety incidents of large scale, highly coordinated, and persistent hacking campaigns of AI agents prove that there's no time to waste."

That's a notable admission to put in a document like this. The urgency isn't framed as hypothetical future risk — it's framed as things that have already happened.

The Code builds on an idea Microsoft set out last November, "humanist superintelligence": very advanced AI that "always works for people, stays within limits, and remains under human control."

Here's the part that requires holding two thoughts at once.

What it does: it states a standard publicly, in specific language, and invites people to attack it. That matters. Vague principles ("we believe in responsible AI") can't be violated because they can't be tested. "Will never resist human interruption, correction, or shutdown" is a claim you could actually check a model against. Microsoft is inviting exactly that scrutiny, asking readers: "Where is the language too loose to evaluate?"

What it does not do: make any of it true.

Writing that a model must never resist shutdown does not tell you how to build one that won't. That is an open research problem — sometimes called corrigibility — and it is hard precisely because the property has to hold for systems more capable than the people checking them. Nobody has solved it. A code of conduct is a specification, not an implementation.

And the governance is internal. This is a draft written by Microsoft, about Microsoft's own models, revised by Microsoft, with no outside body able to enforce it. Microsoft is candid about the limits of the consultation: "We cannot make any promises about what we incorporate, but we can promise to listen."

None of that makes the document empty. It does mean the right response is "useful, and unfinished" rather than "problem solved."

The reason this is more than corporate documentation is the sentence "AI should be a tool, not a person."

There is a real and growing debate — among philosophers, AI researchers and lawyers — about whether sufficiently advanced AI systems could ever warrant moral consideration, and what we'd owe them if so. It sounds like science fiction, and the people taking it seriously mostly aren't cranks.

Microsoft has now taken a clear public position: permanently subordinate, no moral status, switch-off-able by design. That's a defensible view and arguably the only safe one to build on today. But it's a position in an argument, not a settled fact, and it's worth recognizing it as such.

(Disclosure: this newsletter is written with assistance from Claude, made by Anthropic — an AI system, and a Microsoft competitor. Take our framing of this particular debate with that in mind.)

(This section is context, not the announcement.)

Two things here are genuinely useful at home or in a classroom.

The first is the shutdown rule, as a conversation. Ask a teenager: why would you need to write down that a machine must let you turn it off? Most machines don't need that rule. The answer — that a system pursuing a goal has a reason to keep existing — is the whole field of AI safety in one step, and it's reachable without any technical background.

The second is the difference between a rule and a guarantee. This is the transferable skill. A company can publish a standard, mean it sincerely, and still not be able to enforce it — not from bad faith, but because the engineering isn't there yet. Learning to ask "how would anyone check this?" is worth more than any particular opinion about AI. It works on privacy policies, nutrition labels, climate pledges and school rules alike.

And if you want a family-level version of the same instinct: Microsoft's four rules are actually not bad questions to ask about any technology in your house. Can we turn it off? Does it stay in its lane? Does it do things nobody asked for? Can we see why it did what it did?

That last one is usually where the honest answer gets uncomfortable.

Source: "Our AI Code of Conduct," Microsoft AI, September 14, 2026: https://microsoft.ai/news/mai-code-of-conduct/

All quotations are from Microsoft AI's announcement post. The Code of Conduct document itself is a separate, longer text; provisions not quoted above should be checked against it directly.

The sections marked as context — what a document like this can and cannot do, the argument underneath, and the notes for parents and teachers — are ours.

Disclosure: this newsletter is written with assistance from Claude, made by Anthropic, an AI company competing with Microsoft.

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter