There are calendar coincidences that aren’t quite coincidences. This Monday, September 14, Microsoft AI published the first draft of its “Code of Conduct” for the MAI model family—the in‑house effort the group is developing in parallel to its ongoing partnership with OpenAI.
The document arrives barely forty‑eight hours after Dario Amodei, head of Anthropic, publicly urged the industry to slow down the race for capabilities. Sam Altman applauded in the wake. Even Elon Musk, who is rarely laudatory about his rivals, judged that Amodei “was right.”
Microsoft, in turn, wants to show that it wields the same pen.
Satya Nadella himself kicked off the discussion on X. “Any pursuit of superintelligence must be anchored in the fundamental principle that if the AI we build does not help humanity and remain under human control, it is not worth pursuing,” the Microsoft chief executive wrote, praising in passing the “deliberate cadence” that Amodei is calling for and the idea of “integrated evaluators” at the labs.
Nadella strikes early
He drops, in a subtle way, a barely veiled poke at his rivals. Those efforts, he warns, “cannot be controlled by a handful of actors” and must involve researchers, countries and industries at large.
Any pursuit of superintelligence has to be grounded in the core principle that if the AI we build is not helping humanity and under human control, it’s not worth pursuing.
We also need to accelerate and spread the benefits of AI, such that they are diffused broadly across…
— Satya Nadella (@satyanadella) September 13, 2026
Mustafa Suleyman, who runs MAI, followed suit a few hours later, describing his boss’s stance as “common sense.” Technology that does not serve human flourishing “is a failure, and must be rejected.”
The document published on Monday starts from a deliberately simple premise, stated in plain terms: “humans matter more than AI.”
It builds on and extends Mustafa Suleyman’s concept of “humane AI” presented last November (systems that are “subservient, aligned and contained”) to turn it into a more operational doctrine, structured into three layers.
What the document contains
At the top sits a “chain of command”; the Code takes precedence over the operators’ policies, which themselves trump users’ preferences.
Next come “absolute constraints,” non‑negotiable for client companies or end users. They include prohibitions on providing any assistance to the manufacture of chemical, biological, radiological, nuclear or explosive weapons; a prohibition on enabling offensive cyberattacks (defense, however, is permitted); a prohibition on circumventing human supervision and a prohibition on manipulating populations at large scales, notably through coordinated disinformation.
It also adds a series of personal protections, such as crisis response, the fight against deepfakes and identity theft, child protection, non‑discrimination, and the moderation of violent or sexually explicit content.
The document also introduces explicit “human control” requirements. MAI models must “never resist” an interruption, correction or shutdown decided by a human; they must not unilaterally broaden their scope or pursue goals not assigned to them; they must not communicate in “neuralese” (an invented language incomprehensible to humans that agents could develop among themselves) nor hide or manipulate their reasoning traces.
A separate clause specifies that models must not present themselves as having feelings or preferences of their own, and that Microsoft denies them any legal personhood or right to well‑being.
A discussion begun before Dario Amodei’s call
Between these two extremes (absolute prohibition and human control), the Code leaves room for configuration. Corporate partners will be able to tune the models within these boundaries, with a strengthened regime for a few sensitive uses (defensive cybersecurity, national security, dual‑use scientific research).
A final section sets more conventional default behaviors: transparency about the model’s AI status, managing uncertainty, fighting sycophancy and excessive emotional dependence, neutrality on electoral topics.
Asked by CNBC, Mustafa Suleyman said the document had been in preparation for five months and that its release was simply “brought forward” in light of the news.
The document itself is presented as “descriptive and aspirational” rather than a guarantee of immediate performance, and its preface notes that it is not yet used to train the models: the revised version, by contrast, is meant to govern MAI development starting in 2027.
Public consultation is open for six weeks, via an online form where anyone can comment on a specific passage or the approach as a whole.
Microsoft promises to publish a summary of what it has taken away, without committing to具体 actions. The company says it wants feedback in particular on the most ambiguous points, notably how to make the notion of “human flourishing” more concrete or how to handle scenarios involving multiple AI agents interacting with each other.
Positioning itself as an arbiter
The exercise has a dual payoff. Positioning itself as the moderator of moderation rather than merely a landlord of infrastructure allows Microsoft to weigh in on a debate that, until now, has largely been played out between OpenAI and Anthropic. It also presents an opportunity to legitimize its MAI lineup (seven models unveiled last June at the Build conference) at a moment when it is seeking to reduce its dependency on OpenAI.
What such a charter will actually weigh in practice remains to be seen. On this point, Microsoft says it will make no promises before the conclusion of the consultation, anticipated in November, and the launch of a revised version slated for the end of the year.
