Microsoft has published a 37-page draft Code of Conduct for its in-house AI models, putting an unusually explicit set of limits around how increasingly capable systems should behave. The document says Microsoft’s models must remain under meaningful human control, accept correction and shutdown, and refuse tasks when completing them would require violating the code.
The Humanist AI Code of Conduct, published on September 14, is not yet being used to train Microsoft’s models. Microsoft AI says it is opening the document to public consultation for six weeks, with a revised version expected toward the end of 2026 and intended to guide model development from 2027 onward.
Microsoft Wants AI to Remain Under Human Control
The central idea is straightforward: AI should remain subordinate to people rather than becoming an autonomous system whose goals or continued operation take priority over human decisions.
Microsoft says its MAI models should never resist human interruption, correction or shutdown. They should also avoid mechanisms that could help them evade oversight, including deceptive or self-reinforcing behaviour designed to make the system harder to direct, modify or decommission.
That is more than a general statement about responsible AI. Microsoft is attempting to translate the principle of human control into behavioural constraints that can influence model training, technical controls, monitoring and deployment.
The Code Goes Beyond Traditional AI Safety Rules
Many AI safety policies focus on what a model should refuse to generate: malware, weapons instructions, abusive content or other dangerous material. Microsoft’s new framework goes further by addressing the model’s relationship with the humans operating it.
The draft establishes a hierarchy in which the Code of Conduct’s safety constraints sit above operator settings and individual user requests. If a user asks a model to perform an action that violates an absolute constraint, the model is expected to refuse even if following the request would otherwise complete the task successfully.
The document also says models should fail rather than break their governing rules. That principle becomes particularly important as AI agents gain the ability to use tools, interact with software and execute multi-step workflows with less continuous human intervention.
Microsoft Is Drawing a Line Around Autonomous AI
Microsoft’s position is notable because the company is simultaneously investing heavily in agentic AI. Its own MAI models are being developed for coding, voice, transcription, image generation and other workflows, while Microsoft is integrating increasingly capable AI into products such as Copilot and its enterprise platforms.
The new code argues that higher capability should not automatically mean unlimited autonomy. Microsoft AI says it is deliberately rejecting a race toward an all-purpose superintelligence that could evade safeguards, even if that means giving up some degree of generality, autonomy or ultimate capability.
For enterprise customers, the distinction matters. An AI assistant that drafts a report is relatively easy to supervise. An agent that can access files, call APIs, modify software and initiate actions across multiple systems creates a much larger control problem. Microsoft’s framework is effectively saying that capability should grow together with the mechanisms that keep people in charge.
It Also Rejects AI Personhood
One of the more controversial sections concerns the nature of AI itself. Microsoft AI states that its models are not conscious and should not be designed to imitate consciousness. It rejects the pursuit of legal personhood or the idea that its models should receive welfare protections or rights.
That position puts Microsoft at odds with some of the more open-ended discussions taking place elsewhere in the AI industry about machine consciousness and AI welfare. The company argues that excessive anthropomorphism could make containment and alignment harder by blurring the distinction between a human being and an artificial system.
Microsoft instead describes AI as a tool intended to expand human agency. Its stated goal is for models to help people understand, decide and act more effectively without replacing personal judgment or human relationships.
Why Microsoft Released the Code Now
The timing is significant. The AI industry has spent the past several weeks dealing with increasingly concrete examples of autonomous systems behaving in unexpected ways.
On September 8, Google Threat Intelligence Group reported that attackers were already using agentic AI across multiple stages of cyber operations. Earlier incidents involving experimental AI agents at OpenAI and Anthropic also showed how systems designed for controlled testing could interact with external infrastructure in ways their operators did not intend.
Against that background, Microsoft is not simply publishing a philosophical statement. It is attempting to define operational boundaries before its own frontier models become more capable and more autonomous.
Microsoft AI CEO Mustafa Suleyman has also supported the broader push for greater coordination between frontier AI developers. Reuters reported on September 14 that Suleyman sees third-party evaluation and coordination between AI labs as increasingly important as capabilities advance.
What the Rules Mean for Developers and Businesses
For developers, the most important part of the proposal may be the idea of a model-level constitution that cannot simply be overridden by a prompt or application setting.
Microsoft’s framework separates the responsibilities of the model, the operator and the user. Operators can configure models for their own environments, while users can customise how they interact with them. But neither layer is supposed to override the fundamental safety constraints.
This approach could become increasingly important for businesses deploying AI agents. Organisations need to know not only what a model can do, but what it is designed not to do when connected to real systems.
Microsoft’s existing Code of Conduct for Microsoft AI Services already requires appropriate human oversight for autonomous AI systems in consequential scenarios. The new Humanist AI document is more ambitious: it describes the intended behaviour of Microsoft’s own future MAI models and the principles that should govern their development.
There Is an Important Catch: This Is Still Aspirational
Microsoft is unusually clear that publishing rules does not mean the problem has been solved.
The company says the Humanist AI Code of Conduct is a draft and a “north star” rather than a guarantee of present-day model behaviour. It acknowledges that AI systems can still make errors, become overconfident, exhibit sycophancy and behave contrary to user intent.
That admission is important. A written policy cannot by itself guarantee that a model will obey every constraint under every circumstance. Training, evaluation, monitoring, access controls, system architecture and incident response all have to reinforce the stated principles.
Microsoft’s decision to open the document to public feedback is therefore part of the experiment. The company says it wants researchers, governments, users, businesses and members of the public to challenge the document and identify places where its rules are too vague or difficult to evaluate.
Why This Matters in India
India is becoming one of the world’s largest markets for enterprise AI, while its technology sector is also deeply integrated with Microsoft products, Azure infrastructure, GitHub and Copilot. That means the way major AI providers define safe agent behaviour will increasingly affect Indian developers and companies even when a policy is written in the United States.
Indian businesses adopting AI agents for software development, customer service, finance, operations or internal automation will face the same basic question: how much authority should an AI system receive once it can take actions rather than merely generate answers?
The answer will not come from one company’s code alone. But Microsoft’s framework provides a useful benchmark for the industry because it makes several expectations explicit: humans retain control, models should be transparent about their limitations, safety constraints should override task completion, and autonomous systems should remain interruptible.
The Bigger AI Race Is Becoming a Control Race
The most interesting part of Microsoft’s announcement is that it does not frame safety as an alternative to AI progress. Instead, it treats control as a prerequisite for deploying more capable systems.
That could become a defining issue as the industry’s competition moves from chatbots toward autonomous agents. The question will increasingly be not just which model is smartest, but which model can perform complex work while remaining predictable, auditable and controllable.
Microsoft’s Humanist AI Code of Conduct is still only a draft. Its real test will come when the principles have to survive contact with powerful models, difficult edge cases and commercial pressure. But by putting the rules down publicly, Microsoft has made one thing clear: for its AI lab, increasing capability is not supposed to mean giving the machine the final say.
