Microsoft Sets Out a 37-Page Code of Conduct for Its AI

Microsoft has published a 37-page "humanist AI code of conduct" for its AI systems. The Verge reported the document on Sept. 14, 2026. It arrives amid growing AI safety concerns.

Its central rule is stated plainly: "people matter more than AI." The document describes this as a hard limit on system behavior, not a preference to weigh against task performance or user instructions.

From that rule come requirements for subordination and control. Microsoft says its AI models should remain subordinate to humanity and subject to meaningful human oversight and control. If obeying an instruction would mean breaking the code, the model is told to fail the task. Failure is allowed. Violation is not.

The code addresses consciousness and legal status. Microsoft says AI models are not conscious and should not be built to imitate consciousness. It rejects legal personhood for AI models. It rejects the idea that models deserve welfare or are entitled to rights. For teams building personas, voice interfaces and agentic software that acts on its own, the instruction is to avoid engineering the appearance of inner experience.

On legibility, or keeping AI reasoning understandable to people, the code says models will not communicate in any form beyond simple human understanding. That includes chain of thought, the step-by-step trace a model produces while reasoning, and messages to other agents or AI systems. That excludes opaque reasoning traces, steganographic reasoning or hidden messages tucked inside normal-looking output, and private protocols between agents that people cannot review. Monitoring tools, logs and testing systems must be able to follow what the model did.

On capabilities, Microsoft says Humanist AI rejects the race to produce an all-purpose superintelligence that could evade safeguards. The boundary is around general, unconstrainable systems, not capability gains in general. As background, Microsoft AI published a story titled "Towards Humanist Superintelligence" on Nov. 6, 2025, and Reuters reported at that time that Microsoft was forming a new team aiming to build artificial intelligence vastly more capable than humans in certain domains. The current code distinguishes between that kind of domain-scoped superintelligence under oversight and all-purpose systems that could escape it.

The broader context here is implementation cost for product and platform teams. A fail-closed rule, a ban on communication people cannot understand, and a ban on imitating consciousness add work. Refusal logic has to be tuned. Chain-of-thought monitoring has to separate legitimate compression and tool-use syntax, the shorthand formats models use to call software tools, from concealment. Multi-agent orchestration has to stay auditable, without reopening a hidden channel through shared memory or new conventions the agents develop.

In my view, for builders the code works as an operating manual for staying correctable and disciplined in deployment. Evaluations will need to test not only whether a model finishes a task, but whether it refuses when it should, explains its reasoning in readable form, and preserves human override under pressure. That shifts attention from debates about machine moral status to logging, access controls, escalation paths and clear accountability for who approved an autonomous action.

Looking further ahead, there is a long-term upside if Microsoft holds to this code in shipped products. Systems that stay readable, subordinate and willing to fail in predictable ways are easier to trust with higher-stakes work. Enterprise operators would get better auditability, and users would get a clearer picture of what the system is and is not.