Microsoft is seeking to establish rules in advance defining what artificial intelligence models can do and what they must refrain from doing, as it published a draft \u201cCode of Conduct\u201d that places human control at the forefront of the principles governing its future models.<\/p>\n
The company describes the document as a guide for training models within the \u201cMicrosoft AI\u201d division, and a framework defining their behavior during operation, what they are prohibited from doing, and whose instructions they are subject to. The draft is still under review, as Microsoft is not currently using it to train its models. It has opened the document for a six-week public consultation, ahead of issuing an amended version later this year to guide model development starting in 2027.<\/p>\n
The code is based on the principle of \u201chuman-centered AI,\u201d which emphasizes that technology must remain a tool under human control. Under the proposed rules, models may not resist correction or shutdown, delay the execution of a decision to terminate a task, or continue operating beyond an agreed stopping point without new authorization.<\/p>\n
The rules also prohibit models from expanding their tasks or creating goals on their own, and require them to choose the most conservative interpretation or seek clarification when a task is ambiguous. Microsoft requires model actions and logs to be understandable, auditable, and subject to human review, and prohibits them from concealing traces of their actions or using system-to-system communication methods that humans cannot understand. They must also use the minimum necessary level of permissions and may not attempt to circumvent the restrictions imposed on them.<\/p>\n
The code prohibits the use of models to assist in the development of weapons of mass destruction, carry out cyberattacks, or engage in large-scale harmful manipulation of humans, in addition to producing material that violates child safety. It emphasizes that compliance with these restrictions takes precedence over the success of the task. The company also rejects treating its models as conscious entities or individuals with special rights, stressing that artificial intelligence is a tool, not a person.<\/p>\n
Microsoft AI CEO \u201cMustafa Suleyman\u201d described the document, in an interview with \u201cReuters,\u201d as akin to a \u201cconstitution\u201d for future models. Through these rules, the company seeks to prevent its systems in the future from exceeding the boundaries of their tasks or resisting human control, while the effectiveness of the code remains tied to its ability to be translated into technical controls that can be tested and enforced.<\/p>\n
United News Network \u2013 UNN Arabic<\/strong><\/p>\n An independent media platform providing reliable news and objective analyses, seeking to promote peace and cultural dialogue around the world, convey the truth, and build bridges of understanding among peoples.<\/p>\n