Microsoft has presented concrete behavioral rules for its AI models for the first time on Monday and has opened them for public discussion. The code prohibits the systems, among other things, from refusing to shut down or hiding their reasoning from human oversight. Until the end of the six-week consultation, any interested person can submit proposals.
Code Draws Clear Boundaries for Control and Safety
In the Code of Conduct, Microsoft specifies what its so-called MAI models – the company’s proprietary AI systems – must never do. This includes supporting chemical, biological, radiological, or nuclear weapons, offensive cyberattacks, and mass manipulation of people through disinformation. Deepfakes, depictions of child abuse, discrimination, and unlawful surveillance are also included in the absolute prohibitions.
Additionally, the code requires that the models remain under human control at all times. They must not resist interruption, correction, or shutdown, may not unilaterally expand their own scope of tasks, and must not pursue goals that have not been assigned to them. Furthermore, Microsoft prohibits the systems from obscuring their reasoning or traces of actions from human reviewers – a point that is considered sensitive in security research because covert conclusions complicate subsequent oversight. According to Microsoft, the fundamental formula of the document is that people count more than AI: the technology should remain a tool and never claim an independent position.
Debate on Development Pace Shapes the Timing
The code is part of the industry-wide debate about the pace of development of AI systems, which was triggered by Anthropic CEO Dario Amodei with his essay for a conscious speed limit; OpenAI CEO Sam Altman and Tesla founder Elon Musk publicly agreed within hours. Nadella himself had already announced a code of conduct for Microsoft’s models during the market turbulence of the previous day, which beckmann.ai reported – the document was presented on Monday.
On X, Nadella announced the step the day before and wrote that any pursuit of superintelligence must follow the principle that the developed AI serves humanity and remains under human control. Microsoft’s AI chief Mustafa Suleyman stated to IBTimes that the rules had been in the works for five months; however, the growing security debate in the industry influenced the timing of the publication. Involved stakeholders reportedly desired an even more explicit commitment that AI should serve humans and not replace them.
Consultation Runs for Six Weeks Without Firm Commitments
The published text is a draft: until the end of the six-week period, the public can submit feedback via an online form. Microsoft announces that it will evaluate the contributions and publish a summary of the changes but makes no commitments regarding which proposals will actually be incorporated. A revised version of the code is expected to follow later in 2026, but the company has not yet provided a specific date.
According to the company, discussions with academic teams, business partners, and public bodies preceded the draft. The code initially applies only to Microsoft’s own MAI models and not to the additional models from other providers like OpenAI used in Copilot. For companies using Microsoft’s AI systems, there will be no immediate changes to contract terms or functionality due to the consultation – the text describes development principles, not new product features.
It will be crucial whether voluntary codes become binding technical controls – it remains open how Microsoft intends to check compliance should a model actually attempt to evade a shutdown. It is also unclear whether competitors like Google or Meta will follow suit with their own regulations. The revised version after the end of the consultation in six weeks is likely to show how seriously Microsoft takes its own guidelines.
