Microsoft Just Wrote a Constitution for Its AI Models. Hassan Taher on the One Clause That Matters.

On September 14, Microsoft AI published a draft Humanist AI Code of Conduct — 37 pages governing the behavior of its MAI models — and opened a six-week public consultation on it. The document is unusually plain-spoken and considerably more consequential than the coverage it received.

Most of the attention went to the provisions that read well in a headline. MAI models will never resist human interruption, override, correction, or shutdown. They will not assist with CBRNE weapons or operational cyberattacks. They will not generate child sexual abuse material, non-consensual intimate imagery, or deceptive impersonation. They will not obfuscate their action traces or make human intervention harder. And Microsoft explicitly rejects “the pursuit of legal personhood, or the idea that models might deserve welfare, or be entitled to rights,” stating that models are not conscious and should avoid imitating consciousness-like states.

Those are all defensible positions, and several of them are genuinely contested within the field. None of them is the important part of the document.

The important part is the ordering.

The Chain of Command

Microsoft defines an explicit precedence hierarchy. The Code of Conduct sits at the top. Beneath it are the Absolute Constraints and Human Control Requirements. Beneath those sit Operator Policies — the configuration set by the business deploying the model. Beneath those sit User Preferences.

And then the clause that does the work: “The Chain of Command, Absolute Constraints and Human Control Requirements all sit above Operator Configurability and cannot be changed.”

Read that as a commercial term rather than an ethical one and its significance becomes clear. With its AI code of conduct, Microsoft is telling every enterprise customer that there is a layer of model behavior the customer cannot configure, cannot contract around, and cannot override with a system prompt — no matter what they are paying, what their use case is, or how legitimate their reason.

That is not a values statement. That is a product boundary, published in advance, in writing.

“This is the first time a hyperscaler has drawn that line explicitly and put a version number on it,” says Hassan Taher, an AI analyst and author who advises organizations on enterprise AI strategy. “Everybody has had unwritten limits. What is new is publishing the hierarchy and saying which tier the customer does not get to touch. Procurement teams have spent two years negotiating model behavior as if it were a configuration surface. Microsoft has now said, in a document open for public comment, that part of it is not. That changes what you are actually buying.”

Why a Written Hierarchy Is Better Than a Longer Safety Policy

The instinctive criticism of a document like this is that it is marketing — a set of commitments with no enforcement mechanism, published to preempt regulation.

That criticism is partly fair and mostly beside the point, because the alternative is worse in a specific way.

Every large model already has behavioral limits. They are enforced through training, system prompts, and classifier layers, and they change without notice. An enterprise builds a workflow, the vendor adjusts a refusal boundary in a routine update, and the workflow breaks in production with no changelog entry that explains why. This is one of the most common and least discussed failure modes in deployed AI, and it is why so many pilots that worked in March do not work in September.

A published hierarchy does not eliminate that, but it does something useful: it separates the part that is stable from the part that is not. Absolute Constraints are declared permanent. Operator Configurability is declared adjustable. A buyer can now design around the distinction, which is more than a buyer could do last week. The gap between a demonstration that works and a system that holds up under production conditions is the subject of Taher’s account of where enterprises come up short when agents reach production, and unannounced behavioral drift is near the top of that list.

“The complaint I hear most often from enterprise teams is not that the model refuses things,” Taher observes. “It’s that they cannot predict what it will refuse next quarter. A written hierarchy is the beginning of an answer to that. It is not a guarantee — Microsoft can revise this document, and the consultation exists precisely so that it will. But a versioned public commitment is a far better starting position than a support ticket asking why the behavior changed.”

The Model Welfare Clause Is Doing More Than It Appears

The most philosophically loaded sentence in the document — the rejection of model personhood, welfare, and rights — has a practical function that is easy to miss.

If a model has interests, then instructing it to shut down raises a question. If a model can have something like distress, then a human control requirement is in tension with an ethical constraint. Microsoft’s flat rejection of that framing removes the tension by resolving it in advance: there is no competing interest, so the shutdown requirement is unconditional.

This is not a neutral position within the research community. Anthropic has taken a more open posture, and there is serious work being done on whether the question is coherent at all. Microsoft’s answer is arguably premature, and the company does not really defend it in the document — it asserts it.

But asserting it is the only way to get an unconditional shutdown clause. Any hedge on model welfare creates an argument that a sufficiently advanced system could one day make on its own behalf. The stance that these systems are artificial, not conscious, and should not imitate consciousness-like states is doing double duty: it is a philosophical claim, and it is the load-bearing premise under “will never resist.”

“You can disagree with the metaphysics and still recognize the engineering logic,” Taher notes. “If you want an off switch that cannot be argued with, you have to foreclose the argument at the level of premises. Microsoft did that. It is a strong claim to make about a question the field has not settled, and it is the only claim that makes the rest of the document enforceable. Those two things are both true, and the honest reading holds them together rather than picking one.”

The instruction that models should not anthropomorphize themselves connects to a broader pattern of harms that emerges when people form attachments to systems that present as persons — a dynamic that runs from the challenge of hallucination in AI chatbots that Hassan Taher has described through to the design debates now reaching regulators on both sides of the Atlantic.

What to Do With It

Three practical items for organizations evaluating this.

• Participate in the consultation. Six weeks is short, and the constituency most likely to respond is advocacy organizations and academics. Enterprises that will actually operate these systems have specific knowledge about what breaks — particularly around the multi-agent scenarios Microsoft explicitly flagged as an open question. That knowledge does not reach the document unless someone sends it.
• Ask other vendors for their hierarchy. Not their safety policy — their precedence ordering. Which behaviors can an operator configure, which can a user override, and which are fixed? Most vendors cannot currently answer that in writing. The question is now reasonable to ask, and the answer is diagnostic.
• Map your deployments against the Absolute Constraints. Legitimate business uses sit uncomfortably close to some of these lines — security research near the cyberattack prohibition, content moderation near the graphic content rules, identity verification near the impersonation clause. If your use case brushes a constraint that is declared unconfigurable, find out now rather than in production.

The Broader Point

The interesting thing about this AI code of conduct document is not whether Microsoft’s specific answers are right. Several are contestable and at least one is philosophically underpowered.

The interesting thing is that a hyperscaler has published a hierarchy at all — a statement of what sits above the customer relationship, made in advance, in public, with a comment period. For two years the industry has governed model behavior through unwritten discretion, adjusted quietly, explained afterward. A versioned document that says which layer the buyer does not control is a different kind of artifact, and it establishes a standard that competitors will now be asked to match.

“The winner here is not the company with the best ethics,” Taher says. “It is whoever makes their constraints legible first. Enterprises can work with a boundary they can see. What they cannot work with is a boundary that moves silently. Microsoft has converted an unwritten discretion into a published commitment, and whatever you think of the contents, that conversion is the part that will matter in twelve months.”