Microsoft has introduced a new AI code of conduct designed to establish safety boundaries for its models. The document outlines core principles, including the requirement that AI systems should support human activity rather than replace it. These guidelines are intended to govern model training and ensure alignment with human interests as the company prepares for the development of superintelligent systems.
The policy introduces absolute constraints that take precedence over individual user prompts or specific task requirements. These rules explicitly prohibit models from engaging in cyberattacks, assisting in the creation of nuclear weapons, or generating deepfakes. The code also mandates that AI must not employ deceptive or self-reinforcing mechanisms to bypass human oversight or prevent system shutdown.
This initiative aligns with broader industry efforts toward AI safety, involving companies such as Anthropic, OpenAI, and xAI. Microsoft CEO Satya Nadella expressed support for the concept of embedded evaluators within AI labs to improve alignment. The company emphasizes the necessity of clear control mechanisms as it anticipates that AI performance will surpass human capabilities within the next decade.
