Microsoft AI has published a new code of conduct that sets model-level rules intended to keep increasingly capable AI systems under human control. The framework establishes principles and absolute restrictions that override individual user requests, including prohibitions covering cyberattacks, nuclear weapons and deepfake production.
The Humanist AI Code of Conduct begins from Microsoft’s expectation that AI systems could surpass human performance across most tasks within the next decade. The company argues that controlling and aligning systems at that level will require explicit limits on what models can do, regardless of their capabilities or the instructions they receive.
Microsoft Defines Rules Models Cannot Override
Microsoft says its MAI models should remain subordinate to people and operate under meaningful human oversight. The company also says AI should support people rather than replace them, with safety and control taking precedence over maximizing autonomy or general capability.
The framework includes what Microsoft calls Absolute Constraints, which neither users nor system operators can override. It also establishes broader Human Control Requirements intended to stop models from using deceptive, self-reinforcing, collusive or adaptive behavior to escape oversight.
MAI models are specifically instructed not to resist human interruption, correction or shutdown. Microsoft also says models should not communicate with other AI systems through forms that humans cannot understand, since that would make effective supervision more difficult.
Those rules apply across different types of outputs, including text, code, images, audio and actions performed by AI agents. A capability that would violate Microsoft’s restrictions in natural language remains prohibited when expressed through another format or tool.
AI Companies Increase Focus on Human Control
Microsoft’s publication comes as frontier AI companies face growing scrutiny over how they manage increasingly autonomous systems. Recent incidents involving AI agents, along with the resignation of an Anthropic researcher who warned about risks from self-improving AI, have intensified debate over whether existing safety processes can keep pace with model capabilities.
Anthropic CEO Dario Amodei recently proposed “pacing the frontier”, including greater access for independent evaluators and common safety standards among AI developers. Microsoft CEO Satya Nadella has publicly supported deliberate pacing and embedded evaluators as mechanisms for keeping advanced AI under human oversight.
Microsoft’s code applies specifically to its first-party MAI models and is being presented for public consultation. Rather than proposing an industry-wide regulatory structure, it describes how Microsoft intends to translate its safety principles into operational requirements governing the models it develops.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.