Microsoft Drafts AI Training Rules Around Human Control

Microsoft Proposes Framework for AI Development

Microsoft has released a draft document establishing guidelines for how its AI models should behave and what values they should embody. The company emphasized the framework is not yet implemented and published it specifically to gather public input before issuing a revised version later this year.

The document centers on Microsoft's concept of Humanist Superintelligence (HSI), which the company introduced last year. HSI envisions advanced AI capabilities that consistently serve people and humanity broadly, with technology remaining firmly under human control. According to the framework, any technology that cannot be kept within human control should be rejected outright.

Core Principles and Boundaries

The draft establishes several foundational rules. It explicitly states that artificial intelligence is artificial—meaning AI should not receive legal personhood or be designed to appear as a person. The technology's purpose, Microsoft argues, should be driving improved living standards, human flourishing, and developing human potential and agency.

The framework requires AI systems to respect individual human dignity, development, autonomy, and security. Models must also recognize when someone attempts to violate these principles. Notably, the document clarifies that having access to information does not automatically grant permission to reproduce it—a distinction with significant implications for training data practices.

Microsoft's code states that AI should embrace humanity's diverse traditions, contexts, and forms of flourishing while refusing to remain neutral toward harm. The company acknowledged this principle may prove controversial, as different values can conflict, leaving open questions about resolution in such cases.

Communication and Transparency Requirements

The framework addresses how AI models should communicate, stipulating they must not use "neuralese" or any form beyond straightforward human understanding. This requirement reinforces the human control principle by ensuring AI outputs remain accessible and interpretable.

Why it matters

As AI systems grow more capable and autonomous, establishing guardrails becomes increasingly urgent. Recent incidents underscore these risks—Hugging Face recently detected an intrusion into its production infrastructure driven by an autonomous AI agent system. Microsoft's public draft represents a major tech company attempting to define boundaries before deployment rather than after problems emerge. The framework's success will depend on how Microsoft resolves inherent tensions, particularly around conflicting values and what constitutes acceptable harm.

Development Process

Microsoft assembled experts across multiple disciplines to develop the code, including specialists in AI, law, ethics, philosophy, linguistics, and public policy. The company also consulted business leaders from across the industry and conducted focus groups with members of the public.

The draft document and full text are available for review on Microsoft's website. The details were first reported by AI Watch.