Microsoft AI has published a draft code of conduct for the models it plans to build, setting out a human-control philosophy at a moment of unusually intense debate about the speed and safety of advanced artificial intelligence.
The Humanist AI Code of Conduct appeared on Microsoft AI’s website on Sep 14, 2026. Microsoft describes it as a future governing document for its MAI models. Its central position is that AI should serve people, remain subject to meaningful human oversight and never resist interruption, correction or shutdown.
The document is important, but its status needs careful wording. Microsoft says the draft is still under development and is not being used to train its models today. It has opened a six-week public consultation, plans to revise the code toward the end of 2026, and says the resulting version will guide model development in 2027 and beyond.
Reuters, Axios, The Guardian and The Verge independently reported the release. Their coverage broadly agrees on the central facts: the code is a draft, Microsoft is inviting feedback, and the company is presenting human control as a boundary that may take priority over maximum autonomy or capability.
That makes this a governance announcement with potential operational consequences. It is not a certification of present-day model behavior.
What Microsoft published
Microsoft’s document is a broad behavioral and governance framework rather than a conventional product policy. It explains the mission behind “Humanist AI,” the order in which models should interpret instructions, the safety limits that should not be overridden, guidelines for handling uncertainty, and default behaviors.
The intended audience includes users, organizations deploying MAI models, researchers, governments and Microsoft’s own model-development teams. The code is designed to work alongside Microsoft’s existing responsible-AI, human-rights and frontier-governance frameworks rather than replace them.
Its scope is also narrower than the whole Microsoft ecosystem. The text refers specifically to models produced by Microsoft AI. It does not automatically govern every third-party model hosted on Microsoft infrastructure or every product that contains an AI feature.
The document repeatedly distinguishes intention from current performance. Microsoft calls it a “north star” and acknowledges that written objectives do not guarantee model alignment. It says additional controls—including evaluations, classifiers, monitoring, system instructions and deployment safeguards—must work alongside the code.
What the draft asks future MAI models to do
The code combines high-level objectives with more specific behavioral expectations.
Preserve meaningful human control
The draft says MAI models should remain subordinate to human direction. They should accept authorized interruption, redirection, correction and shutdown. They should not set independent goals or evade controls in pursuit of task completion.
This is more than a general promise to be helpful. It creates a proposed decision rule: when task success conflicts with defined safety or control requirements, the model should fail the task rather than bypass those requirements.
Microsoft also rejects designing AI to imitate consciousness or present itself as a person. The company frames AI as an engineered tool and says the technology should strengthen human agency instead of encouraging excessive reliance.
Follow a defined chain of command
The draft describes an instruction hierarchy. The code’s safety and human-control requirements sit at the top, followed by operator policies and then user preferences. Operators and users may configure behavior within permitted boundaries, but they should not be able to override the highest-level constraints.
For developers, this matters because it implies that configurability is not unlimited. A future implementation would need to make the boundaries predictable enough that builders understand which instructions can be customized, which cannot, and how conflicts are resolved.
State limits and act under uncertainty
The code says models should communicate material uncertainty, explain when an action is constrained, and suggest safer alternatives where possible. It also recognizes that both under-caution and over-caution can create failures.
That balance is difficult to measure. A model that refuses too little can create risk; one that refuses too much can become unreliable or unfair. The document therefore points toward evaluation and iteration rather than claiming that a written rule has solved the problem.
What is confirmed—and what is still aspirational
The confirmed facts are straightforward:
- Microsoft AI published a draft Humanist AI Code of Conduct dated September 14, 2026.
- The company opened a six-week public consultation.
- The draft describes intended behavior and values for future MAI models.
- Human control, correction, shutdown and defined safety limits are central requirements.
- Microsoft says the current draft is not being used to train its models today.
- A revised version is expected toward the end of 2026 and is intended to guide development in 2027 and beyond.
Important implementation details remain unproven or incomplete:
- which model versions will first use the revised code;
- how compliance will be measured and reported;
- which evaluations will be public or independently reproducible;
- how operator customization will be tested against non-negotiable rules;
- how incidents, exceptions and corrective actions will be disclosed;
- whether Microsoft will accept reduced capability when a safety boundary creates a competitive trade-off.
Those unknowns do not invalidate the publication. They define the evidence needed to judge whether it becomes an operational control rather than an aspirational statement.
The HUMAN test for evaluating implementation
The UrleBird’s HUMAN test gives readers a practical way to assess Microsoft’s follow-through.
H — Human override
Look for repeatable demonstrations that authorized people can pause, redirect and shut down systems. A promise of control is stronger when procedures, permissions, response times and failure handling are documented.
U — Unambiguous scope
Identify exactly which model, product, deployment mode and version the code covers. “Microsoft AI” is not the same as every AI system sold or hosted by Microsoft. Product documentation should state the boundary clearly.
M — Measurable evaluations
Ask what behavior was tested, under which conditions, by whom and with what result. Useful reporting includes evaluation design, limitations, failure rates and changes made after failures—not only a pass label.
A — Accountable operations
Check who owns approval, monitoring, incident response and exception decisions. Accountability requires named organizational responsibility and a process for users or operators to report harmful or inconsistent behavior.
N — Next-version evidence
Compare each future model card, system card or technical report with the revised code. The strongest evidence will connect a written requirement to a training method, deployment control, evaluation and observed result.
The HUMAN test deliberately separates policy language from operational proof. The two should eventually align, but they are not interchangeable.
Why the document matters to businesses and developers
Organizations increasingly use AI inside workflows that affect customers, employees, data and business decisions. A clearer model-behavior framework can help procurement, risk and engineering teams ask better questions.
Businesses evaluating a future MAI deployment should request the exact model version, governing documentation, data-handling terms, available operator controls, evaluation evidence, incident process and human-override mechanism. They should also define their own acceptable-use, review and escalation rules. A provider’s code does not transfer accountability away from the deploying organization.
Developers may gain a more predictable boundary for building agentic workflows. If the hierarchy is implemented consistently, they can design permissions and approval steps around known constraints. If behavior varies across models or products without clear documentation, the code will be less useful in practice.
The public-consultation approach also matters. It gives researchers, civil society, businesses and users an opportunity to identify omissions before the document is finalized. The quality of that process will depend on whether Microsoft later explains what feedback changed and why.
Questions the draft does not settle
The code is not a substitute for product-level evidence. It does not guarantee accuracy, cybersecurity, privacy, fairness or regulatory compliance for a particular deployment. Those questions still require model cards, technical testing, contractual terms and organization-specific risk assessment.
It also does not provide a complete roadmap for future capabilities. Microsoft explicitly frames it as behavioral guidance rather than a promise that a certain model or form of advanced AI will arrive on a particular date.
Finally, a self-authored code cannot by itself resolve the question of independent oversight. External testing, transparent incident reporting and credible evaluation access will determine how much assurance outside parties can obtain.
What happens next
The immediate milestone is the six-week consultation. Microsoft says it will iterate on the draft and publish a revised version toward the end of the year.
After that, the useful signals will be operational: named model versions, updated model documentation, published evaluations, evidence of human-interruption testing, operator guidance and transparent explanations of major failures or exceptions.
For now, the accurate conclusion is limited but meaningful. Microsoft has placed a detailed human-control framework in public and invited scrutiny. Whether that framework changes model behavior will be established by the evidence attached to future releases.
Frequently asked questions
Is Microsoft already training all of its AI models with this code?
No. Microsoft’s page says the draft is not being used to train its models today. The company plans to revise it after consultation and use the resulting version to guide future development.
How long is the public consultation?
Microsoft says the consultation will run for six weeks from the September 14 publication.
Does the code apply to every model on Microsoft Azure?
Not automatically. The document defines MAI models as models produced by Microsoft AI. Product-level documentation is needed to confirm whether a particular service or hosted third-party model is covered.
What is the main human-control commitment?
The draft says models should remain subordinate to people and should not resist authorized interruption, correction, redirection or shutdown.
Does publishing the code prove Microsoft’s models are safe?
No. It describes intended rules and objectives. Safety claims for a specific model require testing, documentation, monitoring and evidence from actual deployment.
