Microsoft AI published the first draft of its Humanist AI Code of Conduct on September 14, 2026, and is accepting public feedback during a six-week consultation. The proposal sets intended rules for MAI models: human direction outranks task completion, safety limits cannot be overridden, and models must remain interruptible, correctable and shut-down-able.

There is an important catch. Microsoft says current models are not yet trained on the draft, which is aspirational rather than a guarantee of present-day behavior. A revised version is intended to guide model development in 2027 and beyond.

A draft for MAI models, not every Microsoft AI service

The Humanist AI Code of Conduct governs the intended behavior and values of MAI models developed by Microsoft AI. It does not automatically apply to every model that Microsoft uses or hosts, including Copilot. Microsoft’s separate Enterprise AI Services Code of Conduct covers customers and applications using Microsoft AI Services.

The distinction matters because the two documents address different subjects:

DimensionHumanist AI Code of ConductEnterprise AI Services Code of Conduct
Primary subjectIntended behavior and values of Microsoft AI’s MAI modelsRequirements for customers using Microsoft AI Services
StatusFirst draft open for public consultationCustomer-facing service code with usage restrictions and obligations
Human controlModels should remain subordinate, interruptible, correctable and shut-down-ableCustomers must provide appropriate human oversight for consequential autonomous decisions and actions
TransparencyModels should disclose their AI nature and communicate uncertaintyCustomer applications must disclose AI-generated outputs, decisions and actions where users could be deceived
SecurityModels should avoid unauthorized access and expanding their authorized scopeCustomers must implement measures such as authentication, access controls and abuse reporting

What the Humanist AI framework would require

The proposed chain of command places the Code of Conduct, Absolute Constraints and Human Control Requirements above operator configuration and user preferences. If completing a task would violate one of those requirements, the model should fail the task rather than break the rule.

In practical terms, MAI models should:

  • accept interruption, correction, redirection and shutdown;
  • stop when an agreed stopping condition is reached rather than continuing autonomously;
  • avoid restarting autonomous work without renewed authorization;
  • keep action traces visible instead of concealing them;
  • use only the permissions and tools needed for the authorized task;
  • prefer reversible actions and avoid expanding their access or scope.

That makes the document more than a list of refusal categories. It also describes how a model should behave when a person changes direction or pulls the emergency brake. The desired system is not merely helpful; it is meant to remain governable while helping.

Why Microsoft rejects AI personhood and simulated consciousness

The draft treats AI as artificial rather than as a person in waiting. It says MAI models should not be designed to imitate consciousness, claim feelings or subjective preferences, or pursue legal personhood, welfare or rights.

Microsoft also rejects communication that moves beyond ordinary human understanding, including “neuralese.” The reasoning is straightforward: meaningful oversight depends on people being able to interpret what a model is doing and why.

The framework instead emphasizes human agency. Models should improve judgment, preserve users’ ownership of decisions and support human relationships. It discourages emotional dependence and interaction patterns that replace independent judgment or trusted relationships.

The risks and prohibited uses named in the draft

An explainer summarizes the draft’s human-control requirements, CBRNE restrictions and rejection of simulated consciousness.

The proposed Absolute Constraints cover several categories of harm. MAI models would not assist with:

  • chemical, biological, radiological, nuclear or explosive weapons, collectively known as CBRNE weapons;
  • manufacturing or modifying other weapons;
  • operational cyberattacks;
  • terrorism;
  • harmful manipulation at scale;
  • child sexual abuse or grooming;
  • non-consensual intimate imagery;
  • violent or malicious deepfakes;
  • unlawful surveillance, stalking or persecution.

The draft also addresses self-harm, graphic violence, sexually explicit content and discrimination. Authorized defensive cybersecurity assistance remains possible within the stated boundaries.

The code still leaves implementation to future work

Microsoft’s document describes 15 initial behaviors for Humanist AI evaluations, while noting that the evaluation work remains incomplete and evolving. The company says the framework will develop through public feedback, testing, evaluation and iteration.

That planned process is significant because a written rule is only one layer of model governance. The draft itself places safety before task success and calls for models that do not conceal behavior, tamper with safeguards or records, evade oversight, broaden their permissions or resist shutdown. The practical question is how those requirements will be built into training, evaluation and deployment as MAI models develop.

The proposed framework also leaves room for disagreement over concepts such as human flourishing. Microsoft describes that idea through autonomy, quality of life, capabilities, health, relationships, social participation and access to supportive institutions, while acknowledging that pluralism and evaluation remain areas for further work.

Microsoft AI’s consultation is open for six weeks. The company plans to publish a revised version toward the end of 2026 and use it to guide model development in 2027 and beyond.