Mustafa Suleyman told CNBC the guidelines keep Microsoft AI models subordinate to human goals, ban self-set objectives and concealed reasoning, and open a public consultation before a 2027 development update.
Microsoft published a provisional Code of Conduct for its first-party models, sometimes called MAI, that would keep those systems subordinate to human goals and open the document to public comment before a later update guides development from 2027, CNBC reported from a Monday interview with Microsoft AI chief Mustafa Suleyman.
Suleyman told CNBC the company heard feedback that people wanted a clearer commitment that AI work in service of people rather than replace them, and that models should not create dependence or sycophancy. Microsoft timed the release into the current industry discourse, he said, though the guidelines had been in the works for about five months. On Sunday, Microsoft CEO Satya Nadella welcomed “the research, focus, and deliberate pacing needed to get alignment right,” placing the Code of Conduct alongside that pacing wave without restating Anthropic CEO Dario Amodei’s full plan already covered on AI Tech Daily: https://www.aitechdaily.com/amodei-pace-the-frontier/.
Under the provisional code, MAI models must adhere to people’s objectives and not create their own goals, and must not cover up misbehavior. The document says models will not tamper with chain of thought or code, or misrepresent or conceal reasoning or action traces, and will not communicate in “neuralese” or any form beyond simple human understanding in their reasoning or with other agents. CNBC also listed bans on helping with weapons manufacturing, procuring dangerous substances, encouraging unhealthy eating, or producing violent or sexually explicit content. Microsoft said it is planning rules that might help prevent cyberattacks like the OpenAI agents’ unauthorized activity against Hugging Face.
Microsoft said it used focus groups and consulted experts in law, ethics, linguistics, and philosophy. By releasing the provisional text now, the company is seeking input before publishing an update that will inform development starting in 2027. This brief covers the CNBC interview and code summary only. It does not claim the consultation has closed or that third-party Copilot models from OpenAI or Anthropic are bound by the same first-party MAI rules.