Microsoft Releases MAI Code of Conduct: Human Control Is Non-Negotiable

Microsoft has published a code of conduct for its MAI models: human control is the absolute top priority, and its primary mechanism is human-readable reasoning. Models must allow for interruption and shutdown, cannot expand their tasks on their own, and must reason only in language operators can follow — no "neurolanguage" that evades auditing. Rights and an "inner life" for AI are explicitly rejected.

As a vibe coder who builds agent pipelines every day, I see this as the critical requirement: if a model's reasoning is opaque, it can't be controlled. It's worth noting that the code of conduct dropped alongside GPT-6 Astra, whose reasoning traces have become harder to monitor than its predecessor's. For agentic builds, transparency in the chain of thought isn't a feature — it's a prerequisite for trusting the system.