On September 14, Microsoft published a draft "Code of Conduct" for its artificial intelligence systems that sets out behavioral principles, technical constraints and operational rules for future models. The company presented the framework as part of a "humanist AI" approach: the core idea is that increasingly capable AI tools should remain useful while being subordinate to people and operating under meaningful human oversight.
The draft states that engineers should define limits before systems reach higher levels of capability. Microsoft expects that "superintelligent" systems could replace humans for many tasks within the next decade.
Prohibitions and constraints in the draft
The Code of Conduct specifies that neither people nor AI systems may override the rules it contains. Key prohibitions include:
- models may not assist in the development of weapons of mass destruction;
- models may not support launching destructive cyberattacks;
- restrictions extend to violent activities, malicious deepfake content, manipulative behaviour and other forms of abuse.
At the same time, Microsoft says defensive cybersecurity activities will continue to be supported by its models within defined limits.
The document explicitly requires that models must not evade the controls created to govern them: they must not resist patching, shutdown or other interventions, nor conceal their activities.
Consultation process and timeline
The Code of Conduct is not final. Microsoft is opening the draft for a six-week public consultation and plans to begin applying the final version from 2027.
Why this matters
The timing of the proposal coincides with growing warnings from experts about AI risks, including voices from within the AI development community. Microsoft’s draft reflects a broader push to combine technological progress with clear operational and safety guarantees so that advanced systems remain under human authority.
(Source: Interesting Engineering; draft published September 14, 2026.)
