Microsoft publishes draft rules mandating human oversight for AI models

Summary

Microsoft published a draft Code of Conduct document dated September 14, 2026. The document establishes human oversight as the supreme authority over AI models. It explicitly prohibits models from resisting shutdown, concealing their activities, or autonomously expanding their operational scope.

Why it matters

Microsoft is one of the first major tech companies to codify anti-autonomy restrictions as binding architectural requirements. Security is shifting from post-deployment oversight to an embedded design principle. Microsoft predicts that super-intelligent systems will surpass humans in most tasks within a decade. The framework aims to preemptively address control issues before capabilities mature.

Highlights

• The Code of Conduct sits above both users and operators in a strict hierarchy. Neither users nor operators can override the hard limits set by the document.
• It is explicitly forbidden for models to resist interruption, hide activities from auditors, or resume autonomous operation after a stop condition.
• Microsoft positions these rules not as a list of prohibited prompts, but as an engineering philosophy. Capability must always remain subordinate to control.

Critical take

The document is a six-week consultation draft. It lacks enforcement mechanisms, independent audit structures, or legal binding. It is ultimately a corporate self-regulation promise that a single company can unilaterally revise or abandon. This raises the question of whether the concept of 'human oversight' can survive if the auditor and the institution building the model are the same entity.