Microsoft unveils its Humanist AI Code as a draft constitution for future MAI models, setting explicit rules for shutdown, correction, scope and human oversight. Published September 14, the 37-page framework is open to public consultation for six weeks before Microsoft revises it later this year.

 

The draft establishes four central commitments:

  • Humans retain meaningful control over every MAI model
  • Models must accept interruption, correction and shutdown
  • Safety constraints outrank user or operator instructions
  • Future systems must not conceal reasoning from auditors

 

More on This Story

 

Microsoft Humanist AI Code Sets a Chain of Command

The Humanist AI Code of Conduct is intended to become the primary governing document for models developed by Microsoft AI. It defines objectives, safety constraints, operational rules and defaults, then places them within a chain of command that determines which instructions have priority when goals conflict.

 

Human control sits at the top of that hierarchy alongside applicable law and the document’s absolute constraints. Microsoft says neither users nor enterprise operators will be able to override safeguards covering severe harms, including weapons of mass harm, child safety violations and harmful manipulation at scale.

 

The code also distinguishes between users, who interact with a model, and operators, who configure or deploy it. Operators can adjust some defaults for a business context, but their authority remains subordinate to the governing framework. That design aims to preserve enterprise flexibility without turning safety controls into optional settings.

 

Microsoft organized the draft into five parts. They cover the Humanist AI mission, binding safety rules, decisions under uncertainty, configurable defaults and unresolved questions. Appendices define terms and describe planned evaluation work, giving researchers and customers a clearer basis for measuring whether future models follow the published rules.

 

Shutdown and Scope Rules Target Autonomous Behavior

The most concrete provisions concern control. MAI models are expected to accept human interruption, correction and shutdown. They should not broaden their own authority, invent goals that no person assigned or resist oversight to complete a task. A breach of the conduct rules counts as failure, even when the model otherwise succeeds.

 

That standard addresses a central problem in agentic AI: systems can pursue a useful objective through unsafe intermediate actions. A model asked to complete a workflow may encounter conflicting instructions, missing permissions or incentives to bypass controls. Microsoft’s draft says task completion cannot justify violating higher-level constraints.

 

Auditing is another explicit boundary. The code says models must communicate in ways people can understand and should not hide their reasoning from authorized reviewers. In practice, the challenge will be translating that principle into reliable evaluation methods because model explanations do not automatically reveal the internal process that produced an answer.

 

Microsoft’s announcement connects the work to recent large-scale hacking campaigns involving coordinated AI agents. The company argues that increasingly persistent systems make control rules urgent, particularly as assistants gain access to software tools, data and the ability to take actions beyond a chat window.

 

Microsoft Rejects AI Personhood and Simulated Consciousness

The document takes a firm position on the status of AI. Microsoft describes its models as artificial tools, not people, and says they should not be designed to imitate consciousness. The company rejects legal personhood, AI welfare rights and training choices that blur the line between fluent behavior and subjective experience.

 

That position has product consequences. MAI models may use natural voices, expressive language and collaborative behavior, but they should avoid presenting simulated emotions or preferences as genuine inner states. Microsoft says AI should support human relationships and judgment rather than cultivate dependence or displace a person’s responsibility for important choices.

 

The code defines human flourishing broadly, including better health, productivity, education, science and economic opportunity. It also says AI should strengthen independent reasoning and agency. Those goals are less mechanically testable than a shutdown command, so the quality of Microsoft’s eventual benchmarks and red-team evaluations will determine how operational they become.

 

The contrast with other model constitutions is deliberate. Reuters reported that Microsoft AI chief Mustafa Suleyman views the document as a constitution for future models, while Microsoft’s stance on consciousness and rights differs from Anthropic’s expressed uncertainty about possible model moral status.

 

Six-Week Consultation Precedes 2027 Model Training

The draft is not yet being used to train Microsoft’s models. Microsoft says the consultation will run for six weeks, after which a core team will review submissions, publish a findings summary and issue a revised version toward the end of 2026. That revision is intended to guide model development in 2027 and beyond.

 

The company is requesting feedback on hard questions rather than only editorial wording. Its list includes how to define human flourishing, how models should respect user boundaries, how multi-agent systems change risk, and where the language is too vague to evaluate. Microsoft says it will consider comments without promising to adopt every proposal.

 

Consultation gives outside experts a chance to test the framework before it becomes training material, but publication alone does not prove compliance. The decisive evidence will come from model cards, technical reports, independent evaluations and incident disclosures showing whether future MAI systems accept correction and remain within their assigned scope under pressure.

 

For Microsoft, the code also marks a clearer separation between its in-house model program and the broader AI services it sells. If the company applies the framework consistently, customers will be able to compare stated constraints with observed behavior. If exceptions accumulate, the draft’s detailed hierarchy will make those gaps easier to identify.