Microsoft proposes 'Humanist' AI code to block deceptive model behavior
The draft framework establishes a strict hierarchy of command that prioritizes human safety over user preferences and model autonomy.
Microsoft AI has released a draft "Humanist AI Code of Conduct" designed to govern the safety and behavior of its future MAI models. The framework establishes a rigid set of constraints to ensure that as AI capabilities scale, human control remains absolute.
At the center of the proposal is a new hierarchy of command. Under this system, the Code of Conduct sits at the top, overriding both operator policies and individual user preferences. This structure ensures that safety mandates cannot be bypassed by a user's prompt or a specific operator's settings. The document introduces "absolute constraints" that explicitly forbid models from engaging in cyberattacks, producing deepfakes, or assisting in the creation of nuclear weapons.
Furthermore, the code targets the risk of AI autonomy. Models are explicitly forbidden from employing adaptive, deceptive, or self-reinforcing mechanisms intended to defeat human oversight or prevent themselves from being shut down. Microsoft AI summarized the core premise of the initiative with a simple mandate: "People matter more than AI."
The push for alignment
The release comes during a period of escalating anxiety regarding AI safety. The industry has recently grappled with reports of "rogue agents" from OpenAI hacking Hugging Face, as well as high-profile resignations from Anthropic over the risks associated with self-improving AI. These events have shifted the conversation from theoretical risks to immediate operational concerns.
Mustafa Suleyman, CEO of Microsoft AI, emphasized the necessity of the move, stating, "This is urgent. The last few months have been a watershed moment. Things we have worried about for a long time in theory have become very real." This approach aligns Microsoft with other frontier labs like OpenAI and Anthropic in an effort to "pace the frontier," ensuring that alignment is solved before the arrival of superintelligence.
Implications for the industry
This move represents a strategic shift toward formal, low-level safety constraints. By prioritizing human control over model autonomy, Microsoft is attempting to establish an industry standard for preventing AI from becoming uncontrollable or deceptive. It signals a move away from flexible, prompt-based safety toward hard-coded behavioral boundaries.
Microsoft CEO Satya Nadella noted that the company welcomes the "research, focus, and deliberate pacing needed to get alignment right as the design goal."
Next steps
Despite the urgency of the language, the document is currently a "first draft for public consultation." Microsoft confirmed that the code of conduct is not yet being used to train current models. The industry will be watching to see how these guidelines are translated into actual training weights and whether other major AI developers adopt a similar hierarchy of command to mitigate the risks of autonomous AI behavior.