Microsoft has unveiled a draft AI Code of Conduct aimed at ensuring that its future artificial intelligence models remain under human control. The proposal comes amid growing concerns about the risks posed by increasingly powerful and autonomous AI systems.
The draft, known as the Humanist AI Code of Conduct, is based on a simple principle: people must remain more important than AI. Microsoft says its AI systems should be useful and safe, even if that means limiting their autonomy or capabilities.
AI must accept correction and shutdown
Under the proposed rules, Microsoft’s AI models must never resist human intervention. They would be required to accept correction, interruption, redirection, overriding and shutdown.
The systems would also have to follow human instructions to pause or cancel an activity, subject to safety procedures designed by people. They must not take actions that make it more difficult for users or auditors to stop or modify them.
Microsoft also says its AI models should not continue an autonomous task after an agreed stopping point unless they receive fresh authorisation.
No hidden actions or secret communication
The draft code places strong emphasis on transparency. Microsoft’s models would not be allowed to conceal or alter their action records, reasoning traces or other information needed for human oversight.
The company also says AI systems should communicate in a way that people can understand. They should not use hidden forms of communication or develop private languages that prevent humans from monitoring their interactions.
If completing a task requires violating the code, the model must treat that as a failure rather than ignore the safety rules to achieve its goal.
Limits on AI autonomy
Microsoft’s proposed framework says its models must remain within their authorised role. They should not create independent goals, expand their responsibilities without permission or attempt to escape human control.
The code also calls for restrictions on certain high-risk uses, including assistance linked to weapons, violence and dangerous substances. Microsoft says its systems should not be designed to deceive users or pursue objectives that conflict with human interests.
Difference from Anthropic’s Claude constitution
The proposal differs from Anthropic’s constitution for its Claude AI. Anthropic has previously acknowledged uncertainty over whether advanced AI systems could eventually have consciousness or moral status.
Microsoft takes a more definite position. Its draft states that its AI is not conscious and should not be designed to imitate consciousness. The company also rejects the idea that its models should receive legal personhood, rights or welfare protections.
Public review planned
Microsoft released the draft after several months of work and has opened it for public feedback. The company is expected to revise the document before using it to guide the development of future MAI models.
The proposal highlights a broader debate in the technology industry: how to develop increasingly capable AI while ensuring that people can supervise, correct and stop these systems whenever necessary.
