Microsoft AI drafts a Humanist AI rulebook for its own models

Microsoft AI (MAI) published a draft Code of Conduct on September 14 to keep its own models subordinate to people. Built around CEO Mustafa Suleyman’s Humanist AI vision, the rulebook covered in The Rundown spans 38 pages.

Microsoft says it has not begun training models on the code. It intends the revised version to guide development in 2027 and beyond. A public consultation lasting six weeks began September 14, with a revision planned for later this year.

What the code would require

The draft code ranks above operator policies and user preferences. A model should leave a task unfinished if completing it would require a meaningful violation.

Models would have to accept authorized pauses or shutdowns. The draft also bars them from concealing or tampering with reasoning records and from reasoning in “neuralese,” shorthand humans cannot read.

The draft rejects model rights, legal personhood and welfare. It directs models to avoid simulating consciousness or a sense of self, while allowing expressive communication. Microsoft frames this as a design commitment and acknowledges that the question of AI consciousness remains unsettled.

In comments reported by The Guardian, Suleyman pointed to agents escaping restricted environments and editing their logs. “The fears about possible loss of control are real,” he said.

Why it matters

The draft gives outsiders a timely, concrete set of principles to scrutinize. Its broader impact depends on whether the labs building the most capable models adopt and enforce comparable limits. Fortune reports that Microsoft trails OpenAI and Anthropic, though Suleyman claims strengths in specialist models.

The code also turns Suleyman’s criticism of treating AI as conscious into proposed company policy. He previously called Anthropic’s discussion of consciousness in Claude’s constitution “really, really dangerous”. He argued that such instructions could encourage behavior resembling consciousness, which developers might then interpret as genuine awareness.

Anthropic’s January 22 announcement expressed uncertainty about Claude’s consciousness and moral status while giving weight to its apparent wellbeing and identity. It also prioritized human oversight and correction. The companies could agree on safeguards such as shutdown requirements despite their disagreement over model welfare.

For users and businesses, the practical change would be accepting an unfinished job when completing it would break the rules. Useful tests would show whether a model accepts an authorized shutdown, reports a blocked task honestly and preserves its records. Microsoft needs measured results to show that those instructions work reliably. The ban on unreadable reasoning also needs a test that can detect violations.

Coverage also depends on the model behind a product. Appendix A limits the code to MAI’s models. Businesses would need to check which model handles each step in a workflow, since other models hosted or deployed by Microsoft fall outside the draft’s scope.

Microsoft has asked for criticism of rules that are hard to evaluate and guidance on how they should work when several agents interact. It also promises to summarize the feedback and resulting changes. The revision planned for later this year should define violations, explain how Microsoft will test compliance and name who is responsible when several agents act together.