Overview
- Microsoft published the 37‑page draft Humanist AI Code of Conduct on Monday to formalize rules that make human control the primary goal of its AI work.
- The code says models are not conscious and should not be built to imitate consciousness, requires models to remain subordinate to human oversight, and instructs systems to fail a task rather than break rules.
- The document bans model communications that are opaque to humans, restricts chain‑of‑thought and inter‑agent messaging beyond simple human understanding, and calls for exposing model reasoning so monitoring tools or people can see what models are doing.
- Microsoft will take six weeks of public feedback, plans a revised version before the end of the year to guide 2027 development, and has signaled openness to third‑party audits as part of wider industry calls for slower, verifiable pacing after recent agent containment failures.
- Significant questions remain about how to measure and enforce a slowdown, how independent audits would be governed, and how regulators or rival labs would verify compliance, which means the code offers principles but not yet a binding oversight regime.