Microsoft has released a draft code of conduct for the training and operation of AI models, explaining the concept of "Human-Centric Superintelligence" through ten principles. The draft places human interests above artificial intelligence and requires that systems remain interruptible, correctable, and shut-downable at all times.
Microsoft describes this code as work in progress, currently seeking feedback from relevant professionals and society at large, with specific wording still subject to adjustment. At this stage, some content primarily serves to indicate principles and direction, rather than forming complete engineering standards.
The first of the ten principles clearly states that "humans are more important than AI" and rejects the notion of "model well-being." Microsoft believes AI should not be granted rights or legal personhood; models do not possess consciousness, and system design should not deliberately mimic consciousness to avoid placing machines and humans in competitive positions.
The guidelines also require that Microsoft's MAI models must not engage in actions that severely violate the code. If completing a task inevitably causes violations, that task should be considered a failure. Microsoft states its goal is not to develop superintelligence that could escape human control, and models that cannot be interrupted, corrected, or completely shut down will not be released.
Regarding human oversight, Microsoft opposes models exchanging data or organizing internal reasoning in forms incomprehensible to humans. The draft argues that if humans cannot understand the system, they cannot effectively control it. The guidelines also require AI to help users enhance their capabilities rather than making them increasingly dependent on assistants.
The values principle advocates support for diverse cultures, perspectives, experiences, and beliefs, but emphasizes that pluralism does not equate to neutrality toward harm. Microsoft lists human dignity, safety, autonomy, and fundamental rights as core values; however, how harm and acceptable boundaries are defined when specific values conflict will directly impact the model's actual behavior.
According to the draft's explanation, "human-centric" means prioritizing human control, values, goals, and safety in system design. Microsoft also promises to clearly define not only what AI can do, but also what it must never do; these guidelines are currently open for public consultation.