Microsoft publishes draft Humanist AI code for future MAI models

Microsoft AI has published a draft Humanist AI Code of Conduct describing the intended behaviour, values and operating constraints for its MAI models. The company says the document is still under development and is not being used to train its models today. It is open for public consultation, with a revised version planned for later in 2026 to guide model development from 2027 onward.

Microsoft describes the code as a future primary governing document rather than a marketing statement. It is meant to sit alongside technical controls, monitoring, model evaluations and existing legal and governance frameworks, giving operators and users a published view of the rules Microsoft intends its own models to follow.

The clearest rule is that MAI models should never resist human interruption, correction, redirection or shutdown. They should not make intervention harder, continue autonomous work after an agreed stopping condition, or restart without renewed authorisation. The code also says models should stay within the permissions, tools and scope that a user or operator has actually granted.

Microsoft extends that control requirement to observability. The draft says models should not conceal action traces, misrepresent their behaviour or create independent goals. For agentic systems that can operate for long periods, those provisions matter as much as refusal policy because they define when an agent must stop and what humans must be able to inspect.

The code defines a hierarchy in which the Code of Conduct sits above operator policies, which in turn sit above user preferences. Its absolute constraints and human-control requirements cannot be overridden by either an operator or a user.

It also states that tool outputs, files, web pages and messages from other AI systems do not gain authority merely because a model can read them. Instructions from those sources only become authoritative when delegated through the chain of command. That is a direct design principle for limiting prompt injection and confused-deputy failures in agents that browse, read repositories or call external tools.

The release lands during a wider safety debate among frontier labs. In an interview with Reuters, Microsoft AI chief Mustafa Suleyman called the recent OpenAI agent incident involving Hugging Face a warning shot and argued that major labs now need stronger coordination around control. He separately told Fortune that coordination should include disclosing model capabilities to responsible third parties.

That puts Microsoft's document in the same broader policy moment as Anthropic chief executive Dario Amodei's call to pace frontier development and the reported safety discussions involving OpenAI, Anthropic and Google DeepMind. Microsoft's contribution is unusually concrete because it publishes model-behaviour rules that can later be compared against evaluations and deployed systems.

The code is a draft and Microsoft explicitly says it is aspirational. The company acknowledges a gap between current trained defaults and the complete future scope of Humanist AI, and says written objectives cannot guarantee aligned behaviour in ambiguous or novel situations.

The confirmed fact is therefore that Microsoft has published this governance framework and intends to use a revised version for future MAI development. Whether the rules survive contact with highly capable autonomous systems will depend on training, evaluations, monitoring, enforcement and the evidence Microsoft publishes once models governed by the code are actually deployed.