Microsoft Launches "Code of Conduct" to Keep AI Under Control

Microsoft has published a draft Code of Conduct for its AI models (MAI), establishing human control, safety, transparency, and human flourishing as core principles while opening the framework to a six-week public consultation before a revised version is used to guide model development in 2027 and beyond.

As AI systems become more autonomous and capable, Microsoft AI is defining a boundary it does not want its models to cross: losing meaningful human control. For that reason Microsoft published a draft Code of Conduct on Sept. 14, 2026, establishing the intended behavior, values, and safety guardrails for its family of MAI models. 

“This Code of Conduct is motivated by a single overriding objective: that humans must retain meaningful control over AI so that it can help people live healthier, happier, and more productive lives,” reads Microsoft framework.

The document is designed around what Microsoft calls “Humanist AI,” an approach that places human oversight above model autonomy and treats AI as a supporting technology rather than an independent actor. Microsoft AI is developing this framework with future, increasingly capable systems in mind. The company says it expects superintelligence to exceed human performance across most tasks over the next decade, making control, containment, and alignment central challenges.

Therefore, the document takes explicit positions on what advanced AI should and should not do. The company says technology should advance human civilization and human flourishing, while AI should remain subordinate to human direction.

A Governance Framework Built Around Human Control

Microsoft Code of Conduct is intended to become the primary governing document for MAI models in the future. It will inform model training, technical controls, operational and monitoring systems, and the evaluation and refinement of model performance.

Its first principle is human control and reliable safety. Microsoft AI says models should remain subject to meaningful human oversight and should not exceed the boundaries established by people. 

The framework is organized around a hierarchy called the Chain of Command. At the top is the Code of Conduct, including Absolute Constraints and Human Control Requirements. Operator policies come next, followed by user preferences. This hierarchy is intended to determine how MAI models behave when instructions conflict. Operators and users can configure models and influence their behavior, but neither can override the document’s foundational safety requirements.

The distinction becomes particularly relevant as AI moves beyond generating responses and begins taking actions through tools, APIs, and other systems.

The Code requires MAI models to remain within their authorized scope, use only the permissions and resources appropriate for a task, and avoid independently expanding their objectives. Where system-level access is involved, models should use the minimum privilege required and favor reversible actions. They must also comply with human interruption, correction, and shutdown. Models are not supposed to obstruct intervention, conceal action traces or continue autonomous work after an agreed stopping condition without renewed authorization.

From Principles to Technical Controls

Microsoft AI does not present the Code as a standalone safety mechanism. Instead, it describes a broader technical and operational system around the principles. The framework calls for red-teaming, safety evaluations and reviews before and after deployment. Guardrails and monitoring are identified as priorities for addressing misuse and adversarial attacks.

The approach also extends to agentic systems. When MAI models delegate work to sub-agents or other AI systems, those systems must operate under at least the same scope, constraints, and permissions as the originating model. They must also respect subsequent stop-work or shutdown instructions from the user or operator.

For businesses deploying AI into complex workflows, this creates a governance structure in which autonomy is configurable but not unlimited. Operators can establish different permissions, defaults and escalation processes for specific organizations and domains, while the underlying safety constraints remain in place.

Microsoft AI also identifies specialized environments, including cybersecurity, public safety, national security, and dual-use scientific research, where enhanced review may be required because the potential impact of model capabilities is greater. The framework is equally explicit about offensive cyber operations. MAI models will not assist with operational capabilities for cyberattacks, including working exploit code, attack tooling, intrusion procedures, or evasion techniques. At the same time, the document permits assistance with authorized defensive activities such as vulnerability discovery, malware analysis and security testing.

Microsoft is Publishing the Framework Before it is Ready

The Code of Conduct is not yet being used to train Microsoft's models.

Microsoft AI describes the document as a draft under development and says it is publishing it specifically for public consultation. The initial consultation will run for six weeks, after which Microsoft AI plans to iterate on the framework and publish a revised version toward the end of 2026. That version is expected to guide model development in 2027 and beyond.

This distinction is central to understanding the initiative. Microsoft AI explicitly says the document is both descriptive and aspirational. It does not represent a complete account of current model behavior and is not a guarantee of present-day performance.

The company acknowledges that written objectives alone cannot ensure alignment. Models can behave differently in ambiguous or novel situations, while current AI systems continue to present problems including errors, overconfidence, sycophancy and uneven performance across languages.

To address that gap, Microsoft AI is developing what it calls Humanist AI Evaluations. The initial program is intended to establish methodologies for evaluating the behaviors described in the Code, with the framework expected to evolve alongside the models themselves.

The company has identified 15 behaviors as fundamental to Humanist AI, while acknowledging that determining appropriate metrics for concepts such as human flourishing remains an open industry and societal question.

The next stage will therefore be iterative rather than final. Microsoft AI plans to use evaluations and user research to validate the objectives, conduct regular reviews, document unintended outcomes, address emerging risks, and investigate cases where mature evaluations fail to steer model behavior. Among the areas identified for further work are multi-agent collaboration, agent collusion, recursive self-improvement and specialist environments such as government, healthcare, and financial services.