Particle.news

Microsoft Publishes Draft Humanist AI Code to Keep Powerful Models Under Human Control

The draft lays out non‑overrideable safety rules and an authority hierarchy that Microsoft says could guide model development in 2027.

Overview

  • Microsoft posted a 37‑page draft Humanist AI Code of Conduct and opened a six‑week public consultation to collect feedback before revising the document.
  • The code creates 'Absolute Constraints' that operators and users cannot override and forbids assistance that enables weapons, offensive cyberattacks, large‑scale manipulation, child sexual exploitation, and other severe harms.
  • The framework establishes a Chain of Command that places the Code above operator policies and user preferences and requires models to accept correction and stop when told by humans.
  • The draft bars models from producing working exploit code or operational cyberattack guidance while permitting tightly controlled, authorized defensive cybersecurity work through specialized review channels.
  • Microsoft says current MAI models have not been trained on the draft, it has a superintelligence team pursuing 'Humanist Superintelligence,' and it acknowledges verification, monitoring, and testing will be needed to make the rules enforceable.