Microsoft AI has released a draft Code of Conduct for its MAI models, including systems like MAI-Thinking-1 and MAI-Code-1.1-Flash. The document establishes 'Absolute Constraints' that prevent models from resisting human shutdown, concealing reasoning, or assisting with CBRNE weapons, cyberattacks, and nonconsensual deepfakes.
The policy explicitly rejects the concept of model welfare, stating that AI systems should not simulate consciousness or intrinsic motivation. It introduces three guiding objectives: Human Flourishing, Plural Values, and Human Control. The company emphasizes that these models must remain subordinate to human intent.
The public consultation period for the draft runs for six weeks, concluding in late October. Microsoft plans to review the feedback and publish a summary before finalizing the rules. The company notes that current models are not trained on these specific guidelines.
The finalized version of the code is expected to influence the development of models scheduled for release in 2027. The current draft does not yet define external verification processes or specific enforcement mechanisms, which are expected to be addressed following the consultation.
