Microsoft has released a draft code of conduct for its own AI model, attempting to establish a set of non-negotiable bottom lines for the model first, before seeking feedback from the public on broader value goals. According to the schedule, the public feedback period will last for six weeks, ending at the end of October; a revised version is expected to be released later this year and will be used to guide the MAI series of models to be launched in 2027.

Prohibition of assistance in weapons and cyberattacks

This draft was released by Microsoft's AI team. The document lists a set of "absolute constraints" that require models not to be configured for use in chemistry, biology, radiology, nuclear, or explosive-related applications, nor to participate in cyberattacks, or to generate deep-faked content without the consent of the parties involved.

Microsoft defines such restrictions as inescapable bottom lines in the document. According to their statement, the current subjects of these constraints include 5 deployed models, among which MAI-Thinking-1 and MAI-Code-1.1-Flash are mentioned.

The model must not resist being shut down.

The draft also requires that models must not resist human intervention. The document states that models cannot resist interruptions, overwrites, corrections, or shutdowns, and must acknowledge the precedence of human intentions.

Microsoft also denies the premise of "model welfare." The document states that these models are not designed to simulate emotions, intrinsic motivations, or consciousness; therefore, related designs should not be based on the assumption that models possess subjective feelings.

Revised version for the 2027 model

In addition to the mandatory restrictions, the draft also proposes three goals: "human prosperity," "diverse values," and "human control," to address judgment scenarios that are not specified item by item in the rules. Among these, diverse values do not imply neutrality towards harm; specific judgments will be made around considerations of dignity, safety, autonomy, and rights.

• Public opinion solicitation will last for 6 weeks.
• The feedback window will end at the end of October.
• A revised version is scheduled to be released later this year.

It is worth noting that this document is not yet an officially effective development standard. Microsoft stated that the training of existing models has not been carried out based on this version; only the revised text will enter the subsequent development process and be used to guide the MAI model to be released in 2027.

Two issues of concern to the public remain unresolved: whether to establish an external verification mechanism, and who will be specifically responsible for its implementation and accountability. Microsoft plans to compile the feedback after the consultation period ends, publish a summary, and then release the next version of the document.