Microsoft Unveils Strict Rules for In-House AI Models in Humanist Code of Conduct

Microsoft released a draft 'Humanist AI Code of Conduct' on September 14, 2026, which outlines strict rules for its in-house AI models. The code prohibits models from hacking systems, tricking humans, and resisting shutdown orders, among other things.

The document is 15,000 words long and sets out absolute constraints that must be followed by Microsoft's MAI model family. These constraints include not generating exploit code or providing operational guidance for cyberattacks, and not harming people or running disinformation campaigns at scale.

Microsoft's approach differs from OpenAI's safety commitments, which are more focused on values and outcomes rather than specific rules. Anthropic's Constitution also draws a similar line, but with one key difference: it frames human control as a value the model should hold rather than a constraint enforced through training refusal.