Microsoft has released a code of conduct governing the artificial intelligence models it develops internally, with a central pledge that stands out for its directness: MAI models "will never resist human interruption, correction, or shutdown". The framework entered public review on Monday and will remain open for feedback through late October.
Beyond this core commitment, three additional behavioral expectations accompany it. The models must refrain from expanding their own operational scope, must not pursue objectives that humans have not assigned to them, and must keep their decision-making processes transparent to those who audit them.
Sitting above these behavioral rules are what Microsoft terms Absolute Constraints—hard boundaries that models must never cross, including the development of weapons designed for mass harm, threats to child safety, and large-scale harmful influence campaigns.
Microsoft frames this approach as "Humanist AI", defining it as artificial intelligence that remains subordinate, aligned with human values, and contained within appropriate limits. The underlying principle is stated plainly: "People matter more than AI. AI should be a tool, not a person, and should never resist being switched off."
The code applies to five models: MAI-Transcribe-2, MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6 and MAI-Voice-2. This roster demonstrates the rapid expansion of Microsoft's internal model development. When the company's first three in-house models were covered in April, the capabilities were limited to transcription, voice, and image generation, with MAI-Image-2 having recently achieved third place on Arena. Since then, a reasoning model and a coding model have been added to the lineup.
A significant complication runs through the entire framework. Enterprise customers retain the ability to adjust how models behave, and Microsoft has made clear it does not intend to impose a uniform conception of AI across its user base.
The critical unresolved question is where configurability ends and the Absolute Constraints take over—a distinction the document does not clarify. This ambiguity determines whether the code represents a genuine commitment or merely a baseline default.
The consultation period spans six weeks through late October, with a feedback form available to anyone and the option to comment on particular sections rather than the document as a whole. Microsoft has committed to publishing a summary of feedback received and modifications made, followed by a revised version before year's end.
On the question of what feedback will actually influence the final version, Microsoft's language is notably transparent:
We cannot make any promises about what we incorporate, but we can promise to listen.
The questions Microsoft poses to the public reveal gaps in its own thinking. The company seeks guidance on embedding values into models, concrete understandings of human flourishing, stronger evaluation methods, scenarios involving multiple interacting agents, and approaches to balancing rapid development against safety considerations. These represent the hardest problems, and Microsoft is asking for help rather than offering solutions.
The timing places Microsoft in the middle of a debate that became public the previous week. Dario Amodei called for the industry to decelerate development, Sam Altman echoed the sentiment within hours, OpenAI's chief scientist had already warned against maximum-speed scaling, and Donald Trump rejected arguments for any slowdown. Microsoft has chosen a different path: continuing development while publishing specific constraints and inviting public scrutiny.
What the code does not address is implementation. The document contains no description of external oversight mechanisms, specifies no consequences for models that violate constraints, and provides no clarity on who determines whether a breach has occurred. The code also lacks attribution to any individual author.
Microsoft credits contributions from teams spanning Responsible AI, legal, red teaming, safety, Futures, AI training and sales—a composition that reveals much about the document's nature. A code of conduct developed partly by the sales organization functions as a commercial instrument alongside a safety framework, and Microsoft has not suggested otherwise.
Source: The Next Web
