Microsoft unveils draft Code of Conduct for its AI models

Microsoft has unveiled a draft Code of Conduct intended to establish the principles and values that will guide the development of its future MAI artificial intelligence (AI) models, as part of its commitment to a "humanistic" approach to the technology.

The document is currently open to public consultation and sets out the guidelines the company plans to apply both to the training and operation of its AI systems. Microsoft will gather the feedback it receives before preparing the final version, which is expected to be released by the end of this year.

The company notes that these guidelines do not yet apply to the models currently under development. Their purpose is to establish the expected behavior of models planned for 2027, placing people and society at the center while keeping AI under human supervision and control.

One of the principles included in the draft states that these systems should not be designed as people or seek to reproduce characteristics associated with human beings. Microsoft argues that humanistic AI should be designed to help people rather than replace them, stressing that these systems "are not conscious" and should not be developed with the aim of appearing conscious.

The document also addresses the safety measures that should be incorporated into models from the training stage and throughout deployment. These include the ability of systems to identify a hierarchy of instructions and comply with safety restrictions even when they conflict with direct orders from users or operators.

According to Microsoft, the Code of Conduct is intended to define in a "clear and directive" way what its MAI models can and cannot do. The company nevertheless points out that these guidelines do not replace internal or external processes relating to safety, legal matters or the governance of its artificial intelligence systems.

Microsoft currently has seven MAI models, developed for different areas, including voice processing, image generation, programming and safety.