Microsoft AI code of conduct puts human control over models

Microsoft’s new code of conduct matters because it translates years of high level “responsible AI” talk into a concrete rulebook for how its frontier models are supposed to behave. Explicitly stating that models must not pursue their own goals, must not resist shutdown, and must not hack or deceive creates a reference point for regulators and customers to hold the company to. It also stakes out a philosophical position that AI systems, even if they appear agentic, do not warrant rights, welfare or personhood. That is a direct response to Anthropic’s more open ended work on AI welfare and to growing debates about model consciousness.

Practically, this document will feed back into how engineers design reward models, safety layers and incident response. If enforced, it could slow deployment of features that blur the line between tool and agent, such as self modifying code or persistent autonomous workflows. At the same time, by clarifying what Microsoft will not do, it may give enterprises more confidence to adopt its stack at scale, strengthening its position against OpenAI, Anthropic and Google. The big competitive question is whether this “humanist” framing becomes a de facto industry baseline or whether rival labs differentiate by promising more powerful, less constrained agents.