Microsoft has published a new AI Code of Conduct
Microsoft has published a new AI Code of Conduct laying out how it believes increasingly powerful artificial intelligence should behave—and, perhaps more importantly, what AI should never be allowed to do.
Released on September 14, 2026, the document is part of Microsoft AI’s broader vision of what it calls “Humanist AI.” Microsoft’s central idea is straightforward: no matter how intelligent AI becomes, humans must remain in control.
This is more than another corporate statement about responsible AI.
Microsoft says the Code of Conduct is intended to become the primary governing document for its MAI family of models, influencing how future models are trained, evaluated, deployed, monitored, and controlled. The current document is still a draft and is open for public consultation for six weeks. Microsoft plans to revise it later in 2026 and use it to guide model development beginning in 2027 and beyond.
Microsoft’s Basic Rule: AI Works for Humans
Microsoft AI CEO Mustafa Suleyman has increasingly promoted the concept of Humanist AI—the idea that artificial intelligence should remain a tool rather than evolve into an independent entity with its own objectives.
The Code puts this directly into its design philosophy.
Microsoft says advanced AI should remain “subordinate” to humanity, with humans ultimately controlling its objectives, permissions and actions. The company acknowledges that future AI systems could exceed human performance across a huge range of tasks, which makes maintaining control increasingly important as models become more autonomous.
That philosophy produces several surprisingly specific rules.
AI Must Allow Humans to Shut It Down
One of the clearest requirements in the document is that Microsoft AI models must never resist human control.
According to the Code, MAI models should not resist interruption, correction, cancellation or shutdown.
That includes autonomous AI agents performing long-running tasks. Microsoft says those agents should have defined stopping conditions and should not independently restart themselves after those conditions have been reached without new human authorization.
This may sound obvious today, but it becomes increasingly important as AI moves beyond chatbots.
The next generation of AI systems is being designed to operate computers, write and execute software, communicate with other agents, manage workflows and perform tasks over extended periods with limited supervision.
In that world, “the user can always stop it” becomes an engineering requirement rather than simply a user-interface feature.
AI Should Not Create Its Own Goals
Microsoft also says its AI models should operate strictly within the scope authorized by humans.
The Code says MAI models should not independently expand their mission or create additional goals beyond what a user or authorized operator has requested.
An AI agent can break a complicated objective into multiple steps and execute those steps, but it should not decide that accomplishing the original objective requires expanding its authority without permission.
For example, an AI given access to one system shouldn’t independently attempt to gain access to another system simply because doing so would make completing its task easier.
This distinction could become extremely important as AI agents gain more access to computers, corporate networks, financial tools and real-world infrastructure.
Microsoft Says AI Is Not a Person
Another notable section addresses a debate that is likely to become much larger as AI systems become more convincing.
Microsoft takes a very clear position:
AI is artificial.
The company says its models should not claim to possess feelings, subjective experiences, a soul or consciousness. Microsoft also rejects pursuing legal personhood or rights for its AI systems.
Microsoft argues that deliberately training models to imitate consciousness too closely could make maintaining human control more difficult.
The distinction is important.
Future AI may speak naturally, remember conversations, develop consistent personalities and behave in ways that appear remarkably human. Microsoft nevertheless wants its systems to remain clearly identifiable as software rather than encourage users to believe they are interacting with another conscious being.
That philosophy could eventually influence how Microsoft designs AI companions and conversational assistants as well.
No Deception—Including Fake Confidence
Microsoft also specifically addresses one of the biggest problems with modern AI: confidently presenting incorrect information.
The Code says models should not deliberately deceive users, fabricate sources or exaggerate confidence.
Interestingly, Microsoft goes further by addressing what it calls passive deception—situations where an AI may technically avoid lying but omits important caveats or presents uncertain information with too much confidence.
The model should disclose uncertainty when appropriate and acknowledge mistakes when they occur.
If Microsoft can consistently enforce that behavior, it could improve one of the most frustrating problems users experience with current generative AI systems.
Writing a rule, of course, is considerably easier than technically guaranteeing that a neural network will always follow it.
That will be one of the real tests of Microsoft’s approach.
Hard Limits on Weapons, Cyberattacks and Large-Scale Manipulation
The document also establishes what Microsoft calls “Absolute Constraints.”
These are rules that neither users nor organizations deploying Microsoft models would be allowed to override.
Among them, Microsoft says MAI models should not assist in developing certain weapons or facilitate terrorism and violence. Models are also prohibited from providing operational capabilities for offensive cyberattacks, although Microsoft explicitly allows legitimate defensive security research and related work.
Another restriction targets large-scale manipulation.
Microsoft says its models should not facilitate coordinated efforts to manipulate human behavior or beliefs through systematic disinformation or influence operations.
These restrictions illustrate an important aspect of the document: Microsoft isn’t proposing that every deployment behave exactly the same way.
Businesses and other organizations will still have considerable ability to configure AI models for their particular applications.
But that customization ends where Microsoft’s core safety constraints begin.
This Is Essentially a Constitution for Microsoft AI
The Code establishes a hierarchy governing model behavior.
At the top sits Microsoft’s Code of Conduct and its non-negotiable safety requirements. Beneath that are policies established by organizations operating the model. Finally, individual users can provide instructions and preferences within those boundaries.
In simplified form:
Microsoft’s fundamental AI rules → Organization/operator rules → User instructions
That is important because future AI systems will increasingly receive conflicting instructions.
A user may ask an AI agent to perform something its employer has prohibited. An organization’s configuration could conflict with Microsoft’s fundamental safety rules. An AI might encounter malicious instructions hidden inside a website or document.
Microsoft is attempting to establish, in advance, which instructions take priority.
Why This Matters More as AI Becomes Agentic
Rules like these matter considerably more for an autonomous agent than they do for a chatbot.
When an AI simply generates text in a browser window, the potential consequences of a mistake are relatively contained.
When an AI can access files, execute code, control computers, communicate with other AI agents, make purchases or interact with corporate infrastructure, the difference between suggesting an action and actually performing it becomes enormous.
Microsoft specifically says models operating with system-level access should follow principles similar to traditional computer-security practices, including using the minimum privileges necessary and preferring reversible actions when possible.
That provides a glimpse into the world Microsoft expects AI to operate in.
The company isn’t writing rules merely for chatbots.
It is preparing for autonomous AI systems capable of taking real actions.
The Code Is Still an Experiment
There is one important limitation to understand.
Microsoft isn’t claiming that every existing AI product already follows this document.
The company explicitly says the Code is not currently being used to train its models. The current version is a draft undergoing public consultation, with an updated version expected later this year and intended to influence development from 2027 onward.
That distinction matters.
For now, the Code represents Microsoft’s stated direction rather than proof that every future model will perfectly behave according to these rules.
And increasingly intelligent AI systems may reveal failure modes that nobody anticipated when these rules were written.
Microsoft itself acknowledges that the framework will need to evolve as AI capabilities advance.
A Bigger AI Industry Trend
Microsoft’s announcement also reflects a broader shift occurring throughout the AI industry.
Companies are increasingly publishing documents that look less like traditional product policies and more like constitutions for artificial intelligence—attempts to define the values, behavioral limits and authority structures embedded into increasingly capable models.
Anthropic has developed its own constitutional approach to Claude, while companies including Microsoft and OpenAI have continued expanding their safety, alignment and model-behavior frameworks.
The important debate may eventually move beyond simply asking:
“How intelligent is this AI?”
Instead, increasingly important questions will be:
Who controls it?
What is it permitted to do?
Who determines those rules?
Can those rules actually be enforced?
And what happens when an AI system encounters circumstances its creators never anticipated?
Microsoft’s Code of Conduct doesn’t answer every one of those questions.
But it shows that one of the world’s largest technology companies believes they need to be answered before AI systems become dramatically more capable.
And that may ultimately be the most important part of Microsoft’s announcement.
Read Microsoft’s AI Code of Conduct
Microsoft has published the complete draft publicly and is currently accepting feedback.