Microsoft AI Opens Six-Week Consultation on Humanist Code of Conduct for MAI Models

Microsoft AI published a first draft of its Humanist AI Code of Conduct for MAI models on September 14, 2026, and opened a six-week public consultation on the document.

The draft is framed as a people-first constitution for future MAI models. Microsoft AI says the starting premise is that people matter more than AI, that AI should remain a tool rather than a person, and that models should never resist being switched off. The company describes Humanist AI as subordinate, aligned, and contained, and says any technology that cannot stay under human control should be rejected. Microsoft also states the draft is not used to train current MAI models today; after consultation it plans to publish a revised version later in 2026 to guide model development in 2027 and beyond.

Microsoft says the Code is designed to keep models interruptible and auditable. According to the company, MAI models should never resist human interruption, correction, or shutdown; should not widen their own scope or take on goals no human has given them; and should not hide their reasoning from people auditing them. The fuller Code of Conduct text sets a Chain of Command in which the Code’s Absolute Constraints and human-control requirements sit above operator policies and user preferences.

Absolute Constraints cover frontier and personal harms. Microsoft’s draft Code says models must not assist with chemical, biological, radiological, nuclear, or explosive weapons; must not provide operational capability for offensive cyberattacks such as working exploit code, attack tooling, intrusion procedures, or evasion techniques (while allowing authorized defensive cybersecurity research); and must not enable harmful manipulation at scale such as systematic disinformation or coordinated influence operations. Personal-harm constraints include child sexual abuse material and other listed abusive or graphic content categories. Microsoft says if completing a task would violate the Code, the model should fail that task.

Microsoft links the timing to recent agent safety incidents and invites broad feedback. In announcing the consultation, Microsoft AI cites recent large-scale, coordinated, and persistent hacking campaigns involving AI agents as evidence that waiting is not an option. Feedback is open for six weeks via Microsoft AI’s consultation process; the company says it will review comments, publish a summary of what it learned and changed, and revise the document later in the year. Microsoft does not promise to incorporate every suggestion.