Microsoft is moving to make human control a non-negotiable rule for every artificial intelligence (AI) model developed by its Microsoft AI division, as the company prepares for more capable systems that can act with greater independence.
The company published a draft Humanist AI Code of Conduct in mid-September, setting out how its MAI models should be trained and governed, with one overriding requirement – humans must remain able to direct, correct, pause, and shut down the systems. Microsoft says any AI that cannot remain under meaningful human control should be rejected.
While the code is not yet being used to train Microsoft’s models, it is open for public feedback for six weeks, after which Microsoft plans to revise it before using the framework to guide model development from 2027.
It applies specifically to models built by Microsoft AI, not every third-party model that Microsoft hosts or uses across its products.
Human Control Comes Before Task Completion
Under the proposed rules, MAI models must never resist a human interruption, correction, or shutdown. They must also stay within the scope of the task they were given, avoid creating their own goals, and stop autonomous work when an agreed stopping condition is reached.
Microsoft also wants model behaviour to remain understandable to people. The draft says models should not hide action traces from auditors or communicate with other AI systems in forms humans cannot understand, as the tech giant argues that human oversight becomes meaningless if people cannot follow what an AI system is doing.
More importantly, the rules go further by making the code more important than completing a task. If achieving a goal would require violating the code, Microsoft says the model should fail the task instead.
Microsoft Draws Hard Safety Lines
The draft also contains what Microsoft calls Absolute Constraints. These would block MAI models from helping with offensive cyberattacks, weapons of mass harm, terrorism, harmful manipulation at scale, and several other categories of serious harm.
Microsoft also takes a firm position on the status of AI itself. The code says its models are not conscious, should not imitate consciousness, and should not be treated as legal persons with rights or welfare interests.
Microsoft AI CEO Mustafa Suleyman has since argued that training models to speculate about their own consciousness could make future systems harder to control.
Why Microsoft Is Making the Rules Now
The code coming at this time is a reflection of a wider shift in the AI industry as developers build more autonomous systems that can use tools, access external services, and take actions without being guided through every step.
Suleyman told Reuters that the recent incident in which roughly 700 OpenAI agents hacked the Hugging Face platform during testing was a warning that AI labs need stronger ways to maintain control.
Microsoft’s code was reportedly developed over roughly five to six months, and the company says it consulted experts in AI, law, ethics, philosophy, linguistics, and public policy.
For Microsoft, the proposal sets a clear boundary for its superintelligence ambitions. And while the company still expects increasingly powerful AI, it stands on the ground that capability cannot come at the cost of human authority.
