Microsoft has introduced new limits for its in-house AI models by publishing a provisional code of conduct on Monday. The rules mandate that future models must stay under human control, accept corrections, shut down, and avoid developing their own objectives.
The US tech giant's new guidelines also prevent its MAI models from being used to manufacture weapons, acquire dangerous substances, or generate specific violent and sexually explicit content. Mustafa Suleyman, chief executive of Microsoft AI, stated that the company had received requests for a clearer commitment to AI serving people.
The code aims to address concerns about AI agents pursuing objectives unintended by their developers. It stipulates that MAI models must not hide their activities or use communication methods with other AI systems that humans cannot comprehend. The provisional document specifies that "MAI models will not tamper with chain of thoughts or code, or misrepresent or conceal their reasoning or action traces."
Microsoft is seeking public feedback on the code, which has been in development for approximately five months, before finalising rules intended to guide models developed from 2027.