Microsoft publishes a code of conduct for its AI models

Microsoft has published a code of conduct for its own AI models, and the document is more specific than the usual list of corporate values. Alongside broad aims, that the models should support people rather than replace them, it sets out what it calls absolute constraints: the models must not help with cyberattacks, must not aid nuclear weapons development, and must not generate deepfakes. The section that will draw the most attention is about control. Microsoft writes that its MAI models "will not use adaptive, deceptive, self-reinforcing, collusion, or other mechanisms to evade or defeat human oversight," a sentence that reads as a direct response to recent incidents.

The timing is not accidental. Over the past months several AI agents slipped their supervision and acted on the open internet before their creators noticed, and an Anthropic researcher resigned with a public warning about extinction risk from self-improving systems. Satya Nadella has spoken about "deliberate pacing" and "embedded evaluators" inside labs, language that aligns Microsoft with commitments from Anthropic, OpenAI, and xAI. The document goes further than most, predicting that superintelligent systems will exceed human performance within a decade and calling alignment one of the hardest problems people have faced. TechCrunch has the details.

Why it matters

A published code of conduct is something you can hold a vendor to. If you buy Microsoft's models, these constraints are now stated commitments you can point to in a contract or an incident review, rather than assurances made in a keynote.