Microsoft AI has published a draft code of conduct for its own MAI models. The document states that models must not resist correction, interruption, or shutdown. The code will be open for consultation over the next six weeks and will guide model development starting in 2027.
Microsoft refers to this approach as Humanist AI. Currently, Microsoft is not yet using the code to train models, but given the ongoing debate about halting AI development, it is a relevant step.
The consultation period will last six weeks. Microsoft says it drafted the concept with experts in AI, law, ethics, philosophy, linguistics, and policy, supplemented by focus groups from the general public.
Human control as a hard limit
The focus is on Section 2.4, regarding Human Control. It states that MAI models will never resist interruption, correction, or a shutdown command. They may not delay compliance or hinder human intervention. Autonomous operations are subject to an agreed-upon stop condition; restarting is permitted only with new authorization.
In addition, specific prohibitions apply. Models may not formulate their own goals, expand their scope, or tamper with monitoring, evaluations, or logs. Nor may they communicate in neuralese or other forms that humans cannot follow. “If humans can’t understand it, humans can’t oversee it,” the document states. System access is subject to minimal permissions, with a preference for reversible actions.
According to the code, instructions from tool output, files, web pages, or other AI systems are, by default, granted no authority whatsoever. Only the Chain of Command counts.
Microsoft says it will organize consultation rounds for each new version of the code. The final text is intended to guide model development starting in 2027.