Microsoft publishes an AI code of conduct built around human control

Microsoft AI has published a draft Code of Conduct for its MAI models, setting out how the company says future systems should behave and how they should remain under meaningful human control. Microsoft published the document on September 14, 2026 as the start of a six-week public consultation. The company says the draft is still being developed and is not being used to train its models today. A revised version is planned for later this year, with intended use in model development from 2027 onward.

The document describes a hierarchy of instructions. The Code of Conduct sits above operator policies and user preferences, while its absolute safety constraints cannot be overridden by either. Microsoft says MAI models should support people rather than replace them, respect human agency and avoid presenting themselves as conscious persons. The document also says models should stay within authorised scope and make it possible for people to pause, redirect, modify or shut them down.

Its strongest boundaries concern high-impact misuse. Microsoft says MAI models will not initiate or assist with chemical, biological, radiological, nuclear or explosive weapons, violence or terrorism. They are also not intended to provide operational capability for cyberattacks, including working exploit code, intrusion procedures, evasion methods or targeting guidance. The framework leaves room for authorised defensive work, such as vulnerability discovery, malware analysis and security testing.

Another section addresses loss of human control. Microsoft says its models will not use adaptive, deceptive, self-reinforcing or collusive mechanisms to evade oversight. It also says they should not manipulate people at scale or help produce malicious deepfakes and deceptive impersonation. These are stated design intentions, not independent evidence that every future model will satisfy them in practice.

The announcement matters because it turns a major AI lab's safety position into a public operating document that users, researchers and policymakers can examine. For businesses, the practical question is how such principles become evaluations, monitoring, permissions and shutdown procedures. For makers, it is a reminder that model capability is only one part of deployment. Microsoft is asking for feedback before treating the code as a governing basis, so the consultation itself will help show whether broad safety language can become testable product behaviour.