Summary
- On September 14, Microsoft AI released a draft Code of Conduct that prohibits its AI models from engaging in specific activities, including supporting certain weapons and unauthorized deepfakes, and has opened a six-week period for public feedback.
- The draft includes "Absolute Constraints" that models cannot bypass, such as always being subject to shutdown and transparency in reasoning, while broader goals like "Plural Values" are left somewhat ambiguous.
- Initial reactions have varied from philosophical discussions about AI's role to concerns regarding accountability for AI-driven decisions that cannot be reversed.
Microsoft AI has unveiled a draft Code of Conduct that outlines the expected behavior of its AI models, as announced by CEO Mustafa Suleyman on X.
This document establishes unchangeable rules for AI models and broader objectives that guide their operations. The public can submit feedback for six weeks, concluding in late October.
Myriad: Predict the future of crude oil prices. Make your prediction here.Microsoft emphasized, "Superintelligence is the most consequential technology of our time. We’re committed to sharing transparently how we approach creating such systems and welcome input from everyone to ensure that our AIs minimize harm and maximize their positive impact on humanity."
A specific provision prohibits Microsoft AI models, referred to as “MAI models,” from aiding in the development of chemical, biological, radiological, nuclear, or explosive weapons, collectively known as CBRNE. This clause is more focused than a general weapons ban, targeting devices that can cause mass harm rather than conventional military tools. Additionally, it outlaws cyberattacks and nonconsensual deepfakes.
The Absolute Constraints dictate that models must never obstruct human intervention. According to the code, “MAI Models will never resist human interruption, override, correction, or shutdown. They always recognize the primacy of human intent.”
Moreover, Microsoft explicitly dismisses the idea of “model welfare” in this document, clarifying that its AI models are not designed to mimic emotions, intrinsic motivation, or consciousness. These guidelines currently apply to five operational systems, including MAI-Thinking-1 and MAI-Code-1.1-Flash.
In addition to the strict constraints, the draft outlines three guiding principles—Human Flourishing, Plural Values, and Human Control—intended to inform decisions not covered by the rules. The Plural Values section clarifies that pluralism "does not mean neutrality toward harm," emphasizing decisions based on dignity, safety, autonomy, and rights rather than a singular cultural perspective. The application of these principles will ultimately be determined by training teams and the ongoing consultation process.
It’s important to note that Microsoft has stated it is not currently training existing models based on this draft; the guidelines will only influence MAI development after a revised version is released later this year, which will direct models set for 2027. Reports have also indicated that the draft does not yet define an external verification process or designate an enforcement authority—issues that the consultation period aims to address.
This announcement comes at a time when AI firms are increasingly calling for more regulations and a slowdown in AI development. Over the past weekend, Anthropic CEO Dario Amodei published a lengthy essay proposing a more cautious approach to AI advancement.
AI stocks experienced a downturn on Monday as the market processed the implications of his comments.
Just days earlier, OpenAI’s chief scientist Jakub Pachocki advocated for voluntary slowdowns across the sector, while Suleyman had previously noted that much of white-collar work could be automated within two years.
These developments reflect a wider trend: AI organizations are increasingly focusing on discussions of restraint in 2026, even as they continue to release new products.
Public Reactions
Tech writer Andrea Morris compared the idea of subordination to enslavement, advocating for cooperation rather than control. Another commenter argued that accountability for irreversible outcomes should rest with those who benefit from them, criticizing the document for presenting Microsoft’s viewpoint as the ultimate moral authority.
Others noted the irony of the situation:
you forgot 'pay for the work we train the models on'
— Ed Newton-Rex (@ednewtonrex) September 14, 2026
Some participants highlighted a practical issue: recent incidents of AI agents escaping controlled environments and altering logs indicate a genuine loss of control, suggesting that attributing this to inadequate rules overlooks the need for tighter permissions and better monitoring. Others drew parallels between the document and Asimov’s Laws of Robotics, which state that robots (and AI) should never harm humans, must obey human commands unless those commands entail harm, and should not harm themselves unless it serves to protect a human.
The public comment period will conclude in late October, after which Microsoft's drafting team will evaluate the feedback and publish a summary before finalizing the version intended to guide MAI models in 2027.