Microsoft has released a 37-page draft of its Humanistic AI Code of Conduct, specifically targeting its proprietary MAI models. The core principle establishes that humans hold authority over AI, which is defined strictly as a tool that must never refuse a shutdown command.
The guidelines mandate that MAI models must comply with human interruptions or cancellations without delay. Models are prohibited from restarting autonomously, expanding their own permissions, or pursuing unauthorized objectives. Furthermore, they must not use cryptic communication or conceal their reasoning processes.
The draft outlines strict prohibitions, including assistance with weapons of mass destruction, offensive cyberattacks, and large-scale public opinion manipulation. Microsoft explicitly states that these models lack consciousness and should not be granted legal personhood or rights.
The hierarchy of control prioritizes the code of conduct over operator policies, which in turn override user preferences. Adjustments to models in defense or public safety sectors require mandatory review and authorization from Microsoft.
The proposal is currently open for a six-week public discussion period. Microsoft intends to finalize the rules by the end of the year, with full implementation for the training and development of the MAI series scheduled to begin in 2027.
