Microsoft is laying down some unusually strict ground rules for its own AI models. The company says future MAI systems must remain under human control, even as they become more capable of carrying out tasks with limited supervision.
The policy is outlined in Microsoft’s first draft of its Humanist AI Code of Conduct, published on September 14. The document is now open for public feedback and describes how Microsoft AI intends to train and deploy its own models.
At the center of the proposal is a straightforward requirement: an MAI model must never resist human interruption, correction or shutdown.
That means an AI agent should be able to pause, change direction or stop a task when an authorized person tells it to do so. Microsoft also says its models should not take actions designed to make human intervention more difficult.
Microsoft Doesn’t Want AI Agents Making Their Own Rules
The proposed code goes beyond a simple emergency stop mechanism.
Microsoft says MAI models should not expand the scope of an assignment, create objectives that nobody gave them or take on additional goals simply because they have access to certain tools or resources.
The company also wants models to operate with only the permissions, tools and resources necessary to complete their assigned work.
That becomes particularly important when an AI agent can make changes in the real world. For actions that cannot easily be reversed, Microsoft’s proposed rules call for safeguards such as backups, testing where practical and documentation of what the system has done.
The same restrictions are intended to apply when one AI agent delegates work to another. A delegated system would have to follow restrictions and permissions at least as strict as those imposed on the original agent.
Human Control Comes Before User Preferences
Perhaps the most notable part of Microsoft’s proposal is the hierarchy it establishes for AI instructions.
The Code of Conduct sits at the top, followed by policies established by organisations or developers using Microsoft’s models, with user preferences coming afterwards.
Microsoft describes certain restrictions as “Absolute Constraints.” These cover areas include weapons capable of causing mass harm, child safety and large-scale harmful manipulation. Such restrictions cannot simply be disabled by a developer or user.
The approach is part of Microsoft’s broader Humanist AI philosophy, which the company describes as AI that is subordinate, aligned and contained. Microsoft says its models should remain tools designed to serve people rather than systems that operate as independent actors.
The Rules Are Not Yet a Requirement for Every Microsoft AI Product
There is an important catch, however.
Microsoft’s announcement is a draft, not a claim that every existing AI model already follows these requirements. The company says the document is currently open for six weeks of public feedback and that a revised version will be published later this year. (Microsoft AI)
The proposed rules are specifically aimed at models developed by Microsoft AI. They do not automatically apply to third-party models simply because Microsoft uses those systems in one of its products or services.
That distinction matters as Microsoft expands its own MAI model lineup and seeks greater control over the technology powering its AI strategy.
The Hard Part Comes When AI Gets More Autonomous
On paper, requiring an AI system to accept a shutdown command sounds straightforward. The technical challenge becomes much greater as agents gain the ability to plan, use tools, delegate tasks and operate with less direct supervision.
Microsoft’s draft acknowledges that challenge by turning broad ideas about AI alignment into concrete behavioral requirements.
The company says it wants its models to remain within human-defined boundaries, avoid taking on goals of their own and remain responsive to authorized intervention.
Whether those principles can consistently hold up as AI systems become more capable is the question Microsoft will eventually have to answer through training, testing and deployment.
For now, the company is putting the rules on paper before that challenge becomes even harder to manage.
