Microsoft's New House Rules for AI: People Come First

Microsoft has published a 37-page rulebook for its AI systems. The Verge reported the document on Sept. 14, 2026. It comes amid growing concerns about AI safety.

The main rule is simple: "people matter more than AI." Microsoft says this is a hard limit, not a suggestion. The AI must follow it even if that means doing a task poorly or refusing a user instruction.

The rulebook says AI must stay under human control. People must be able to watch it and step in. If finishing a task would break the rules, the AI should fail the task. It is better to fail than to break the rules.

Microsoft also says its AI is not conscious, and it should not act as if it is. The company rejects legal personhood for AI models. It rejects the idea that models deserve care or have rights. Teams building chat personalities, voice assistants and AI helpers that act on their own should not make them seem to have feelings or inner lives.

The AI must also keep its thinking readable to people, like a student asked to show their work. Microsoft says its models will not use any form of communication beyond simple human understanding. That includes chain of thought, the private notes a model writes while solving a problem, and messages to other agents or AI systems. That rules out hidden reasoning, secret messages tucked inside normal answers, and private languages between AIs that people cannot check. Logs and safety tests must be able to follow what happened.

On power, Microsoft says Humanist AI rejects the race to build an all-purpose superintelligence that could escape safety controls. The limit is on general systems that cannot be controlled, not on smarter AI 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 new code draws a line between that kind of focused superintelligence under oversight and all-purpose systems that could escape it.

The broader context here is extra work for product and platform teams. A fail-rather-than-break rule, a ban on unreadable communication, and a ban on faking consciousness take effort to enforce. Refusal rules must be tuned. Checks on chain of thought must tell normal shorthand for using tools apart from hidden messages. Systems where many AIs work together must stay open to review, without hidden channels through shared memory or new habits the AIs develop.

In my view, for builders this is a practical checklist. Tests must ask not only whether the model finished the job, but whether it said no at the right time, explained itself clearly, and let a human take over under pressure. That means less focus on whether machines have rights, and more focus on logs, access controls, escalation steps and clear accountability for who approved an AI action.

Looking ahead, there is a long-term benefit if Microsoft follows this code in real products. AI that stays readable, stays under control and fails in predictable ways is easier to trust with important jobs. Companies get clearer records, and everyday users get a clearer sense of what the system is and is not.