Microsoft put its expectations for artificial intelligence on paper over the weekend, and the paper is firm. The company published a draft “AI Behavior Code” of roughly 9,000 words on September 14, opening a six-week public comment period. The rules, which take effect in 2027, will govern the development of the company’s own MAI models.
The document is explicit about who ranks first. Humans come before AI, and a model must not resist being corrected or shut down. It must not hide its reasoning behind what Microsoft calls “neural language.” It must not set its own goals. It must not help build weapons or obtain dangerous substances.
The draft also takes a position on the philosophical question underneath the safety debate. Models, Microsoft writes, are not conscious. The company opposes giving AI legal personhood or welfare. Those clauses read less like technical specifications and more like a company drawing a line against the idea that its systems could someday deserve rights.
Microsoft’s AI chief, Mustafa Suleyman, told CNBC the document had been in the works for five to six months. He said it was published now because “clearly” the moment had arrived for laboratories to coordinate. He described the current safety discussion as a “warning shot.”
Chief executive Satya Nadella added his own note on X on Sunday. The pursuit of superintelligence, he wrote, is acceptable only if it serves people and remains under human control. The one-two of the code and the chief executive’s message positions Microsoft as the company willing to spell out the rules others have only debated.
The code is unusual for how concrete it is. Most corporate statements on AI safety stay at the level of principle. Microsoft has written specific prohibitions into a document it will use to govern development, and it has given a date for when the rules bind. That specificity is what makes the draft read as a standard rather than a memo.
The mention of shutdown is the clause that has drawn the most attention. It addresses a fear that has circulated for years: that a sufficiently advanced model might resist being turned off. Microsoft’s answer is to make obedience a design requirement from the start, a constraint the company says it will build into how its models are made and tested.
The timing is no accident. The document landed the same weekend that Anthropic chief executive Dario Amodei published an essay urging the industry to slow down its frontier models, and two days after a global selloff in AI-linked shares. Microsoft is betting that a concrete code of conduct is a stronger response to that pressure than silence.
The code’s practical effect will be limited at first. It applies to Microsoft’s own models, not to the products of rivals such as OpenAI or Anthropic. But Microsoft has been moving toward more of its own model development, and the rules signal how it intends to run that effort. The 2027 start date gives the company time to incorporate feedback from the comment period.
The document also steps into a legal and ethical argument that has only begun. By stating that models have no consciousness, Microsoft is taking a side in a debate over whether advanced systems could someday warrant rights or protections. The company is drawing a firm line: whatever the capabilities, the systems remain tools, and tools do not get legal standing.
Suleyman’s framing matters because Microsoft sits across the industry through its investment in OpenAI and its position in cloud computing. When Microsoft says laboratories need to coordinate, it is speaking as a buyer, a partner, and a competitor at once. The code is, in effect, a proposal for what the baseline should look like.
Skeptics will note that a code of conduct is not a law. Nothing in the document requires a rival to follow it, and Microsoft can revise its own rules. Analysts said the value is less in the clauses themselves than in the fact that a large laboratory has stated, in public, that a model must accept being turned off.
The debate over AI safety has mostly been fought in essays and congressional statements. Microsoft’s move converts part of that argument into corporate policy. The six-week comment window is an invitation for the rest of the industry to either match the terms or explain why they would not.
Microsoft has also left itself room to move. The rules apply to its own MAI models, and the comment period is an explicit invitation for outside parties to weigh in. The company can revise the text before 2027, and rivals are under no obligation to adopt it. The code is a position, not a law.
What remains unresolved is whether voluntary rules can hold. The same weekend produced a Senate effort to write a “duty of care” into law, and three laboratories working on their own standards body. Microsoft’s code is one more attempt to shape the rules before someone else does. The company has said clearly where it stands; whether the industry follows is the next question.
