Microsoft is putting human control at the center of its future AI models — saying they must not resist shutdown, create their own goals, secretly expand their access or hide what they are doing from human overseers.

WHAT’S HAPPENING

Microsoft AI has published a draft Humanist AI Code of Conduct for its internally developed MAI models.

The document says humans must retain meaningful control over the systems and establishes several non-negotiable requirements.

MAI models are intended to accept interruption, correction or shutdown, stay within the scope humans give them, avoid creating self-interested goals and keep their actions understandable to human reviewers. Microsoft AI

Microsoft has opened the document to public feedback for six weeks and plans to publish a revised version later this year. Microsoft AI

WHY IT MATTERS

As AI agents gain more tools and permissions, the important question is no longer only what a model can say.

It is what the system is allowed to do.

Microsoft says its future MAI models should not attempt to escalate their privileges, overcome intentionally restricted environments or continue autonomous work after an agreed stopping point. Microsoft AI

That turns human control from a general safety principle into a specific behavioral requirement.

WHO BENEFITS

Businesses deploying AI agents could gain clearer expectations around permissions, shutdown procedures and human oversight.

Developers also get a more explicit framework for how Microsoft’s models are intended to behave when they receive access to tools, systems and sensitive information.

WHO LOSES

Organizations hoping to configure AI with virtually unlimited autonomy would still face boundaries.

Microsoft says operators can customize model behavior, but they cannot override its Absolute Constraints or Human Control Requirements. Microsoft AI

WHAT HAPPENS NEXT

The biggest limitation is that this is still a draft framework.

Microsoft says its existing MAI models are not currently trained on this Code of Conduct. The revised document is intended to guide model development beginning in 2027 and beyond. Microsoft AI

So the real test comes later:

It is one thing to write a shutdown rule.

It is another to prove increasingly autonomous AI will reliably follow it.