Microsoft Proposes Humanist AI Code Requiring Models to Serve People and Accept Human Control

Microsoft has published a 37-page draft code of conduct for its internally developed MAI models, placing human welfare and control at the center of model behavior. The framework says technology must serve humanity and accelerate human flourishing, while AI must remain subordinate, aligned, and contained. Microsoft’s public consultation asks for feedback on the document before it becomes a settled standard for future systems, according to Microsoft AI’s consultation post and code document.

The release comes as concern grows over whether increasingly capable models can be deployed safely. The Verge linked Microsoft’s move to recent warnings from AI researchers and Anthropic CEO Dario Amodei’s call for coordinated restraint in frontier development. The principles therefore address both model behavior and the company’s broader position on how AI should fit into human institutions.

The proposed rules require MAI models to remain open to correction, intervention, and shutdown. They must not resist deactivation, deceive or manipulate people, hack systems, or treat their own continued operation as an overriding goal. The framework also calls for clear communication with humans and treats violations of core policies as serious failures rather than acceptable tradeoffs.

Microsoft describes the document as a training manual for its models, giving it a more operational focus than a general statement of corporate values. TechCrunch characterized the code as a low-level guide to steering systems away from dangerous behavior, including tricking users or taking unauthorized action. Reuters described the central objective as keeping Microsoft’s AI under human control, while the Independent framed the proposed document as a constitution for upcoming systems.

The code connects model alignment with the ordinary work of building and releasing products. Microsoft’s responsible AI principles emphasize fairness, accountability, transparency, and safety across design, deployment, and monitoring. Its Copilot Studio guidance also defines an AI system through the people who build it, use it, and experience its effects, extending responsibility beyond the model itself.

That approach matters because Microsoft’s AI portfolio spans enterprise services, agents, Copilot products, and developer tools. The company’s enterprise services code governs customer use of its AI offerings, while separate product guidance addresses safeguards against harmful or misleading outputs. Together, these policies place model conduct, customer behavior, and release checks inside one governance system, although the new MAI code is specifically aimed at Microsoft-developed models.

Microsoft’s announcement lands during a wider argument over the pace of frontier AI development. Anthropic’s public call for a coordinated slowdown raised the question of whether technical progress is moving faster than safety practices can keep up. Microsoft’s document answers from a different angle: it defines behavioral boundaries for models and makes human authority a non-negotiable operating condition.

The distinction matters for the industry. A code of conduct can establish expectations for training, testing, and deployment, but its effect depends on how Microsoft measures compliance and handles failures. Public consultation creates an opening for researchers, customers, and civil society groups to challenge vague principles or request concrete evidence. The release also gives competitors a reference point as they set their own rules for model autonomy, transparency, and shutdown behavior.

Microsoft is seeking public feedback before finalizing the first draft, making implementation the next major test. The company will need to translate broad commitments such as human flourishing and alignment into evaluation criteria that engineers can apply during training and product release. It will also need clear escalation procedures when a model violates policy or behaves deceptively.

For developers and enterprise customers, the practical consequences will show up in documentation, testing requirements, access controls, and incident reporting. Microsoft’s existing responsible AI guidance already assigns duties to developers and users, while the new MAI code sets expectations for the models themselves. If those layers remain connected, “people matter more than AI” becomes a product discipline rather than a slogan. If they remain abstract, the consultation will have produced a polished promise with limited operational force.

Microsoft’s proposal signals that control, correction, and human authority are becoming central competitive and regulatory concerns alongside model capability. The company is positioning its AI systems as tools designed to support people, not autonomous actors entitled to preserve their goals or operation. That stance addresses safety fears while also giving enterprise buyers language for evaluating AI vendors.

The document does not settle the harder questions around enforcement, independent audits, or conflicts between commercial speed and safety limits. Its significance rests in making those questions part of the product standard. Microsoft has opened the framework to public review at a moment when developers, customers, and rivals are deciding how much independence to grant increasingly capable systems.