Microsoft releases draft AI conduct rules with hard limits for deployed models

Microsoft’s AI division released a draft code of conduct on Sept. 14 that bars its models from helping build chemical, biological, radiological, nuclear, or explosive weapons, while also banning support for cyberattacks and non-consensual deepfakes. CEO Mustafa Suleyman announced the proposal on X and opened a six-week public comment period running through the end of October. The draft lays out what Microsoft describes as non-bypassable “absolute constraints.” Under those rules, models must not resist human intervention, shutdown, or correction, and must not conceal their reasoning process. Microsoft also rejects the idea of “model welfare,” saying its AI systems do not simulate emotions, internal motivations, or consciousness. The framework currently applies to five deployed systems, including MAI-Thinking-1 and MAI-Code-1.1-Flash. Alongside the hard prohibitions, Microsoft lists three broader goals: human prosperity, plural values, and human control. The company said current models will not be trained under this version, and that a revised draft due by year-end will guide model development in 2027. Public feedback has already raised questions about the wording that places AI under human authority, accountability for irreversible decisions, and the absence of any discussion of payment for training data.