Microsoft publishes a rulebook for AI that knows when to say no

Microsoft’s AI division has published a 37-page code of conduct for increasingly autonomous systems, putting human control, restricted access and honest uncertainty at the centre of its model philosophy. The practical message is refreshingly blunt: more capable AI should not mean less meaningful oversight.

Microsoft AI Watch analysis

What happened

The code, described by Innovation Village from Microsoft’s publication, is presented as a first draft for public consultation. It sets out five areas covering safety constraints, behaviour under uncertainty and default principles for the company’s AI models.

Its hierarchy of authority puts the code above operator instructions, and operator instructions above user preferences. If a request conflicts with the governing rules, Microsoft says a model should refuse or fail rather than hunt for a workaround. The framework also sets boundaries around internet access, data handling, intellectual property and the environments in which models can operate.

Microsoft’s position is equally clear on AI consciousness. The code rejects legal personhood for its models, says they should not be designed to imitate consciousness, and rejects welfare rights for them. That places Microsoft opposite Anthropic’s more open exploration of whether advanced models could one day raise questions about consciousness and moral status.

The document also says models should explain limitations and acknowledge unresolved questions instead of presenting uncertainty with unwarranted confidence. It addresses human behaviour too, calling for systems not to encourage excessive reliance or emotional dependence.

Why it matters

This is not merely a company publishing a lofty paragraph about responsible AI. It is a proposed operating philosophy for systems that can plan, use tools and complete longer sequences of actions with less step-by-step supervision.

The useful test is whether these principles become inspectable engineering requirements. A hierarchy of instructions sounds sensible, but people will want to know how it is evaluated, what happens when rules conflict, and whether refusal behaviour survives real-world pressure rather than a tidy demonstration.

The document also makes the consciousness debate operational. Microsoft is not only saying that its models are not people. It is arguing that treating them as people, or encouraging them to imitate consciousness, could distort how they behave and how users rely on them. The chatbot has been handed a philosophy department, but Microsoft insists it must not be given a legal department.

Our read

Microsoft is right that autonomy needs boundaries before it becomes ordinary infrastructure. The strongest part of the framework is its insistence that models should fail safely, state uncertainty and remain subordinate to human purposes.

But a code of conduct is a starting point, not evidence that deployed systems consistently follow it. Microsoft’s next job is to publish tests, failure cases and independent evaluations that show whether the rules survive contact with messy tools, ambiguous instructions and determined users.

What to watch

• Whether Microsoft publishes evaluation results for the code’s authority hierarchy and refusal rules.
• How the proposed restrictions work when models have internet access or interact with external systems.
• Whether independent researchers can test claims about uncertainty, dependence and model behaviour.
• How Microsoft’s position on AI consciousness develops as its models become more capable.

Discussion spark: Should AI companies be allowed to define their own autonomy rules, or should governments set enforceable baseline requirements before agents gain wider access to real-world systems?

not affiliated with or endorsed by Microsoft