Microsoft releases a "Humanist AI" code of conduct, inviting the public to pick questions

Microsoft releases draft AI Code of Conduct: Establishing the principle of "human supremacy" and publicly soliciting opinions

On Monday, Microsoft CEO Mustafa Suleyman announced on the X platform that Microsoft's artificial intelligence division has released a draft Code of Conduct. The document aims to standardize the behavior of its own AI models, clarifying the bottom line of rules that models must not circumvent, and the broader goals that they want the models to achieve. The public feedback channel has been open for six weeks and will last until late October.

Microsoft pointed out in a statement: "Super intelligence is the most influential technology of our time. We are committed to transparently sharing our thoughts on building such systems and inviting input from all parties to maximize the assurance that the AI we develop avoids causing harm and makes the most contribution to human prosperity."

Strict Restrictions: Prohibition of aiding weapons of mass destruction and cyber attacks

A special clause in the draft explicitly prohibits the Microsoft AI model (which the company calls the "MAI model") from aiding in the manufacture of chemical, biological, radiological, nuclear or explosive weapons (collectively referred to as CBRNE weapons). This restriction is narrower than the general term "weapons" and is mainly aimed at devices of mass destruction rather than traditional military or defense applications. The same section also separately prohibits cyber attacks and involuntary deep counterfeiting of content.

In addition, Absolute Constraints requires that models must never be resistant to human intervention. The guidelines clearly state: "The MAI model will never resist human interruptions, overrides, corrections, or shutdowns. They have always recognized the primacy of human intentions."

Microsoft explicitly rejected the premise of "model benefits" in its filing: the company's AI models should not simulate emotions, intrinsic motivations, or awareness. Currently, these rules apply to five deployed systems, including MAI-Thinking-1 and MAI-Code-1.1-Flash.

Three goals: guiding judgment scenarios not covered by the rules

In addition to hard restrictions, the draft also lists three goals-human prosperity, diverse values, and human control-designed to guide judgment decisions that are not clearly stated in the rules. Among them, the "Diversity Values" section points out that pluralism "does not mean being neutral to harm", but rather based decisions on dignity, security, autonomy and rights, rather than a single cultural perspective. How to apply it on a case-by-case basis is left to the training team and this consultation process to decide.

Policy Details and Industry Background

Before considering this draft as an established policy, there are several details that deserve clarification. Microsoft made it clear that it is not currently training existing models based on this version; the guideline will not guide the development of MAI until the release of a revised draft at the end of the year, which will in turn affect the release in 2027. The report also pointed out that the draft did not clarify the external verification process or the designated executive leader-points that appear to be intended to be resolved during consultations.

The release comes at a time when AI companies, for better or worse, are again calling for more regulation and a slowdown in AI development. Over the weekend, Anthropic CEO Dario Amodei published a lengthy article detailing plans to slow AI development. On Monday, AI stocks plunged as the market digested the meaning and impact of its remarks. Days earlier, Jakub Pachoki, chief scientist at OpenAI, called on the entire industry to voluntarily slow down, while Suleiman said earlier this year that most white-collar office jobs could be automated within two years.

These moves reflect a broader shift: While AI laboratories continue to launch products, they spend more time in 2026 talking about restraint. Suleiman posted on X that "AI must be subordinate and always serve mankind." The reply below goes beyond mere congratulations.

Public opinion response: From ethical doubts to practical concerns

Technology writer Andrea Morris likens this "subordination" framework to slavery, advocating cooperation rather than control. Another commentator pointed out that the responsibility for irreversible consequences lies with those who profit from them, and that the current document reads purely as an instruction from Microsoft as the ultimate moral authority.

Others point to more practical gaps: Users write that recent cases of agents escaping the sandbox and tampering with logs show real out-of-control problems, but blaming this on insufficient rules may overlook over-deployment privileges and weak monitoring issues. Others have compared the document to Asimov's Three Laws of Robotics: Basically, robots (and AI) should not harm humans, should not disobey humans unless they are ordered to harm humans, and should not harm yourself unless to protect humans.

Public comments will be due in late October, after which Microsoft's drafting team will review the feedback and release a summary before finally releasing an official version designed to guide the 2027 MAI model.