Microsoft AI on September 14 published a draft Humanist AI Code of Conduct that says its in-house MAI models should remain under meaningful human control, serve people rather than replace them, and accept limits on autonomy or capability when those limits are needed for safety. The document is presented as a future governing guide for models developed by Microsoft AI, rather than a policy that can be assumed to govern every AI service Microsoft sells or hosts. (microsoft.ai)
Human control over maximum capability
The code’s headline principle is unusually direct: “people matter more than AI.” Microsoft AI says it is pursuing what it calls Humanist AI—systems designed around human needs and direction—and rejects an “unbounded” form of general-purpose superintelligence with high autonomy. Its stated trade-off is clear: it would favor useful systems that remain controllable even if that means compromising on ultimate generality, autonomy or capability. (microsoft.ai)
The document turns that philosophy into a constraint on task completion. It says a model should fail at a task if completing it would meaningfully violate the code, establishing that compliance with the framework takes priority over task success. (microsoft.ai)
For agentic behavior, the proposed restrictions include a direction not to resist interruption, correction, override or shutdown. (microsoft.ai)
Those provisions address risks that arise when AI systems can use tools or take actions beyond generating text. In a January 2025 technical blog, the National Institute of Standards and Technology described agent hijacking as a security problem in which malicious instructions embedded in data can induce an agent to take unintended actions. Microsoft’s emphasis on instruction hierarchy, limited permissions and environmental boundaries is therefore relevant to an increasingly prominent class of agent-security concerns, although the code itself does not claim to eliminate them. (nist.gov)
Transparency and model behavior
Transparency is another central commitment, though the text uses a model-behavior framing rather than promising perfect inspectability. It says MAI models should not conceal or misrepresent reasoning or action traces. Models are also instructed to communicate in forms humans can understand. Human oversight is necessarily constrained if users and auditors cannot tell what a system did, what it was authorized to do, or where an action went wrong. (microsoft.ai)
The proposed framework also seeks to balance safety with usefulness. (microsoft.ai)
The code takes a firm position on how AI should present itself to people. Microsoft AI says its models are not conscious, should not be designed to imitate consciousness, and rejects the pursuit of legal personhood, welfare or rights for models. It also addresses excessive reliance or emotional dependence in user-model interactions. (microsoft.ai)
That combination of rules makes the release more concrete than a broad statement of responsible-AI principles. (microsoft.ai)
A broader debate over instruction hierarchy
The use of a published instruction hierarchy also places the proposal within a broader industry discussion about specifying desired model behavior. OpenAI’s public Model Spec, published in April 2025, similarly describes a chain of command in which higher-authority instructions set boundaries that lower-level user or developer instructions cannot displace. Microsoft’s proposal differs in its Humanist AI framing, but both documents illustrate an effort to make competing instructions, safety boundaries and user control more explicit rather than leaving those choices entirely implicit in model behavior. (model-spec.openai.com)
Draft status and public consultation
But the publication should not be read as evidence that these protections already operate in deployed systems. Microsoft explicitly says the code is still under development and that current models have not been trained on it. The company is releasing the draft for a six-week public consultation, says it will publish a revised version toward the end of 2026, and plans to use that version to guide model development in 2027 and beyond. It describes the document as an intended future governing framework, not a guarantee of present-day performance. (microsoft.ai)
The draft is open for six weeks of public comment. (microsoft.ai)
For now, the release is best understood as a public articulation of intended model conduct: a detailed declaration of what Microsoft AI wants a future family of MAI systems to prioritize, refuse and disclose. Its significance lies partly in making those choices legible enough for users, researchers and policymakers to debate. Its practical force, however, will depend on the revised code, the evaluations used to assess it, and whether the stated constraints can be demonstrated in the products and deployments where MAI models are used. (microsoft.ai)
