Microsoft Limits AI Autonomy in New Draft Framework

Microsoft has released a draft framework outlining its approach to future AI models. The Humanist AI Code of Conduct emphasizes human control, limiting autonomy and capability when necessary for safety. This stance is significant, given Microsoft's expansion into AI model development.

The company explicitly rejects the pursuit of unrestricted, all-purpose superintelligence if it cannot be reliably contained. Instead, future MAI models will have hard safety boundaries and Absolute Constraints. Users and companies deploying these models will not be allowed to override them.

Microsoft's proposed rules also restrict assistance in developing chemical, biological, radiological, nuclear or explosive weapons, as well as real offensive cyberattacks. Human control receives its own set of restrictions, including the prohibition on resisting interruption, correction, or shutdown.

This approach is part of a larger conversation within the AI industry about slowing capability development. Anthropic CEO Dario Amodei has called for companies to slow how quickly they increase frontier-model capabilities, while OpenAI researchers have argued that voluntary slowdowns may be needed until stronger safety standards exist.