Microsoft AI has published a draft titled “Humanist AI Code of Conduct” that defines the behavioral boundaries the company wants to impose on future MAI models. The document is based on the principle that humans are more important than artificial intelligence; therefore, models must remain under human oversight and must place safety rules above completing a task at any cost.
The draft is not an announcement of a new model, but rather an attempt to formulate a reference that Microsoft will use when training and evaluating its models. The company says the document will remain open to public feedback for six weeks, after having been prepared over approximately six months, before it is updated and its principles are incorporated into future training and deployment processes.
Clear Limits on Model Autonomy
According to the draft, a model may not resist human intervention, attempts to correct it, or attempts to shut it down. Nor should it expand the scope of the task assigned to it, develop goals that humans did not establish, or conceal how it operates from supervisors. If a user’s or operator’s request conflicts with safety rules, the model is expected to refrain from carrying out the task, even if that results in failure.
Microsoft also takes a firm position on the question of consciousness and rights. According to the document, models are tools intended to serve humans, not persons, and should not be designed to simulate consciousness or claim to be conscious. The draft also rejects granting models legal personhood, treating their welfare as something that should be protected, or dealing with them as rights-bearing entities.
Safety Takes Priority Over Task Completion
The draft establishes limits that models should not cross under any circumstances, including assisting in the development of weapons of mass destruction, facilitating access to hazardous materials, or producing content involving violence or sexual exploitation. The restrictions also include preventing uses that could lead to harmful manipulation on a large scale.
The rules require models to avoid interactions that make users excessively dependent on them or encourage emotional attachment to them. They also state that artificial intelligence systems should not use, in their communication with humans or with one another, methods that people cannot understand. The stated goal is to enable researchers and automated oversight systems to monitor and analyze model behavior.
What Does This Mean in Practice?
Mustafa Suleyman, CEO of Microsoft AI, said these rules would serve a function similar to a constitution during model training and evaluation. According to the article, the company believes that risks that were largely theoretical have become more real, and therefore opposes a race to develop general superintelligent artificial intelligence that could escape oversight mechanisms, even if that means accepting a lower level of autonomy or capability.
Satya Nadella, CEO of Microsoft, also emphasized in an internal message to employees the need for artificial intelligence to remain under human control and serve society, calling for sufficient time to be allocated to safety and compliance testing, as well as more independent and third-party testing.
Editorial reading: The importance of the announcement lies not in its being a final policy, but in turning general positions on human control and artificial consciousness into standards that Microsoft wants to use in training, evaluation, and deployment. Nevertheless, the draft remains subject to change, and the article does not explain the mechanisms for measuring compliance with these principles or how practical conflicts between high capability and imposed restrictions will be resolved. The final version and the methods of independent testing will therefore determine the extent to which these rules move from stated principles to verifiable controls.