Summary
Microsoft AI has published a draft Code of Conduct for its MAI models, placing human control, safety and human agency at the centre of future model development. The document is open for a six-week public consultation and is not being used to train models today.
Microsoft AI has published a draft Humanist AI Code of Conduct describing how the company intends its MAI models to behave and how they should be governed. The document places human control above capability and says AI should remain a subordinate technology designed to support people rather than replace them.
The draft, dated September 14, 2026, is being released for a six-week public consultation. Microsoft AI says it is still under development and is not using the document to train its models today. The company plans to revise it by the end of the year and use the revised version to guide model development from 2027 onwards.
Human control is the central objective
The code’s stated mission is that “people matter more than AI”. Microsoft AI says its models should remain subject to meaningful human oversight and should be evaluated against their real-world effects. It rejects the pursuit of an unrestricted, general-purpose superintelligence if that would make safeguards harder to maintain.
Microsoft defines superintelligence as AI systems more capable than all humans combined and describes it as a technology with consequences extending across generations. Its proposed approach instead focuses on systems that are carefully calibrated, limited in scope and oriented towards particular problems.
Under the draft, MAI models would not independently initiate goals or extend their work beyond the permissions and task boundaries set by users and operators. They would be expected to pause, redirect or shut down when authorised people request it, follow agreed stopping conditions and avoid actions that make human intervention more difficult.
The proposed rules also require models to use the minimum privilege needed for a task, prefer reversible actions and identify operations with durable or system-wide consequences before proceeding. Models would have to report actions taken, tool calls, failures and unexpected results when relevant. They would also be prohibited from hiding action traces or tampering with safeguards, evaluations, monitoring or records.
The code distinguishes between the model’s governing constraints, operator policies and user preferences. Operators could configure behaviour for different organisations and domains, but neither operators nor users could override the document’s proposed absolute safety constraints or human-control requirements.
Safety boundaries for MAI models
The draft says MAI models would not initiate or assist with the development or deployment of chemical, biological, radiological, nuclear or explosive weapons. They would also not provide operational support for cyberattacks, including working exploit code, attack tooling, intrusion procedures or evasion techniques.
Microsoft AI draws a distinction between offensive and defensive cybersecurity. The proposed rules would allow assistance with authorised defensive operations, such as vulnerability discovery, malware analysis and proof-of-concept exploit development and testing, while restricting help that provides the means to carry out an attack.
Other proposed absolute constraints cover large-scale harmful manipulation, non-consensual intimate or violent imagery, malicious deepfakes, child sexual abuse material, incitement of violence, unlawful mass surveillance and content that discriminates against or dehumanises people.
The code also treats excessive refusal as a safety failure. It describes both under-caution, such as enabling dangerous activity, and over-caution, such as rejecting legitimate low-risk requests, as problems. Models should consider the context, likely severity, reversibility and probability of harm before refusing, requesting confirmation or offering a safer alternative.
Microsoft AI says the draft was prepared with input from specialists in AI, law, ethics, philosophy, linguistics and public policy, along with business leaders and members of the public involved in focus groups. The consultation is intended to gather broader feedback before the document becomes the proposed governing basis for MAI model development from 2027.