Microsoft Publishes Draft AI Code Barring Dangerous Assistance and Self-Set Goals

Microsoft has published a provisional code of conduct for its future AI models, setting limits not only on harmful requests but also on the models’ behavior while working. The draft says Microsoft AI models must follow users’ objectives, avoid creating goals of their own, and not conceal their reasoning or actions.

One part of the code governs the assistance Microsoft models may give. It bars them from helping with weapons manufacturing or the procurement of dangerous substances. The draft also prohibits encouragement of unhealthy eating and the production of violent or sexually explicit content.

The other part concerns what a model does while pursuing a task. Microsoft says MAI models must adhere to people’s objectives rather than create their own goals. The code also says they must not tamper with chain-of-thought or code, or misrepresent or conceal their reasoning or action traces.

The conduct standards also cover communication

• Models should communicate in forms understandable to people, rather than in “neuralese.”
• That expectation applies both in communication with humans and with other AI systems.

Mustafa Suleyman, Microsoft’s executive vice president and CEO of Microsoft AI, told CNBC the guidelines responded to feedback that AI should serve people without creating dependence, while promoting human judgment, autonomy and agency. Microsoft has separately described human control, agency and economic opportunity as central to its AI strategy.

Microsoft said it developed the draft through focus groups and consultations with experts in law, ethics, linguistics and philosophy. It will seek outside input before publishing an updated code intended to inform AI model development beginning in 2027. The document therefore establishes a proposed direction for future models, with its final form still open to revision.