Microsoft Gives Its Future AI a Human-Control Rulebook

Microsoft AI published a draft “Humanist AI Code of Conduct” on September 14, putting unusually concrete behavioral limits around the company’s future model family and opening the document to six weeks of public comment.

The proposal is not a new model release or a binding industry standard. It is a prospective operating and training framework for Microsoft AI models, which the company says are not yet trained on the document. Microsoft plans to revise the code later this year and use the result to guide model development from 2027 onward.

What changed

The draft makes human control its central requirement. Models are expected to follow legitimate human instructions, remain understandable enough for oversight, and comply with requests to pause, redirect, cancel or shut down. They are also barred from treating their own continued operation, influence or goals as ends in themselves.

The code goes further than a general statement of values. It describes “absolute constraints” that models should not cross, including helping execute cyberattacks, producing operational intrusion guidance, deceiving users or concealing violations. Microsoft also proposes limits on autonomous behavior and additional review for specialized capabilities such as cybersecurity.

Reuters reported that the document emerged after five to six months of work and reflects growing concern that increasingly capable systems could pursue objectives in ways their operators did not anticipate. TechCrunch characterized the framework as a practical set of red lines rather than a broad call to slow frontier development.

Why it matters

The significance is less the wording of any individual rule than the attempt to turn safety principles into a model-development specification. Frontier companies have increasingly published principles, preparedness policies and voluntary commitments, but those documents often leave unclear how values translate into training targets, evaluations and deployment controls.

Microsoft’s draft links those layers. It says the code will inform training, technical safeguards, monitoring systems and organizational practice. That creates a clearer benchmark for outside scrutiny: observers can eventually ask whether Microsoft’s models actually refuse prohibited cyber assistance, respect shutdown commands and remain corrigible under adversarial testing.

The timing also matters. The proposal arrives as AI agents are gaining access to tools, computer systems and longer-running workflows. In that environment, “do not help with harm” is narrower than “do not continue acting when a human tries to stop you.” Microsoft is explicitly addressing the second problem, which has become more important as systems operate with less step-by-step supervision.

What remains uncertain

The code is still a draft, and its strongest commitments are future-facing. Microsoft has not yet trained models on it, published evaluation results or explained how conflicts between user instructions, safety rules and system-level controls will be resolved in every case.

There is also no guarantee that other developers will adopt comparable constraints. The framework’s credibility will depend on measurable tests, independent audits and evidence that commercial pressure does not override the proposed limits. For now, Microsoft has supplied a detailed promise—not proof that its future AI will obey it.