Key facts
- Microsoft AI released a draft Code of Conduct for its AI models on September 14.
- The code prohibits AI models from assisting with CBRNE weapons, cyberattacks, and nonconsensual deepfakes.
- AI models must never resist human intervention, override, correction, or shutdown.
- Microsoft AI models are not intended to simulate feelings, intrinsic motivation, or consciousness.
- The draft code is open for public feedback for six weeks, through late October.
- Microsoft plans to release a revised draft toward the end of the year, guiding 2027 releases.
Microsoft AI has released a draft Code of Conduct for its AI models, setting forth principles and constraints for their behavior. The document, published on September 14, aims to ensure AI systems remain under human control, never resist shutdown, and communicate understandably. It prohibits AI models from assisting with chemical, biological, radiological, nuclear, or explosive weapons (CBRNE), cyberattacks, and nonconsensual deepfakes.
Microsoft AI CEO Mustafa Suleyman announced the draft on X, describing it as a "constitution" for future AI models and emphasizing that AI must be subordinate and always in service of people. The code includes "Absolute Constraints" that models cannot be configured around, such as never resisting human intervention or concealing their reasoning. It also explicitly rejects the concept of "model welfare," meaning the AI models are not meant to simulate feelings or consciousness.
The company is seeking public feedback on the draft for six weeks, through late October. This initiative follows recent calls from leaders at AI companies like Anthropic and OpenAI for a more measured pace of AI development. Microsoft stated that the current models are not trained on this version of the code, and a revised draft is expected toward the end of the year to guide 2027 releases.