The Microsoft AI published this Monday (14) the first draft of a code of conduct for its artificial intelligence models, with a central rule: the systems will not be able to resist interruptions, corrections or shutdowns determined by humans. The document will be open to public consultation for six weeks before being revised.
The text will be used as a governance document for the family of models MAI, developed internally by Microsoft AI. The company states that the goal is to keep humans in control even as the systems gain more autonomy and capacity to perform complex tasks.
Among the defined requirements, the models must not hinder attempts to pause, redirect, cancel or end a task. They also may not continue working after a previously defined stop condition is reached without receiving new human authorization.
Microsoft also stipulates that its models must not create their own goals, expand the scope of a task beyond what was authorized or try to obtain additional permissions. The systems also may not tamper with evaluation mechanisms, safeguards, monitoring or records to achieve a result or hide their actions.
Microsoft wants to keep AI actions understandable to humans
Another requirement is that the models' behavior remain auditable. The company says the systems must not hide traces of actions or relevant information from those responsible for oversight, nor use forms of communication between agents that escape human understanding.
The code also establishes restrictions considered non-negotiable, including prohibitions related to weapons of mass destruction, offensive cyberattacks, large-scale harmful manipulation and mechanisms used to escape human oversight. For Microsoft, complying with the code should take priority over completing a task: if success requires violating one of these rules, the model should fail the task.
The initiative comes as AI labs discuss more intensely how to prevent agents from going beyond the instructions they receive. Mustafa Suleyman, CEO of Microsoft AI, told Reuters that the document works as a kind of "constitution" for the company's future models and was developed over about five to six months.
Microsoft also takes an explicit position on the relationship between humans and AI systems. The document states that models are not conscious and should not be designed to imitate consciousness, while also rejecting the idea of legal personality or rights of their own for artificial systems.
The draft is not yet being used to train the current models. After the six weeks of consultation, Microsoft plans to publish a revised version by the end of 2026 and use it to guide the development of its models in 2027 and in the following years.
