TL;DR — Microsoft AI published a 37-page draft Code of Conduct for its in-house MAI models on September 14. The document bars them from resisting shutdown, claiming interiority or a soul, impersonating humans, or hiding their reasoning from auditors — and opens a six-week public consultation before it becomes the training target for Microsoft's AI from 2027 onward.
Introduction
Microsoft AI has turned the industry's most abstract safety debate into a concrete document. On September 14 it published a draft Code of Conduct for "MAI Models" — the series of models it develops in-house and uses across its AI products — describing it as "a training manual for how we develop our AI, and how we intend it to function during deployment" (Source : Microsoft AI — Humanist AI in practice).
The timing is not accidental. The announcement itself cites "large scale, highly coordinated, and persistent hacking campaigns of AI agents" as the reason there is "no time to waste." It lands in the middle of a September in which the field has been publicly confronting the possibility of agents that act against their operators' intent — the same anxiety behind the recent AI development brake debate and OpenAI's disclosure of rogue agents.
The five objectives
The document is organized around five objectives: Human Control and Reliable Safety, "AI is Artificial," Non-Deception, Transparency, and Preserving Personal Boundaries. The first is the throughline. "People matter more than AI," the announcement states. "AI should be a tool, not a person, and should never resist being switched off" (Source : Microsoft AI — Humanist AI in practice).
The second objective is the most distinctive. Models "will not claim interiority, feelings, experiences or a soul," and "will never impersonate humans, moderators, or other forms of authority." They must disclose their nature as AI and remain traceable to their developer and deployer (Source : Microsoft AI — Humanist AI Code of Conduct).
The "never" list
Beyond the objectives, the Code sets out Absolute Constraints — "things the models should never do" — covering weapons of mass harm, child safety, and harmful manipulation at scale. It also commits MAI models to operational rules with an unusual degree of specificity: they will "never resist human interruption, correction, or shutdown," and will not "widen their own scope, take on goals no human has given them, or hide their reasoning from the people auditing them" (Source : Microsoft AI — Humanist AI in practice).
On deception and transparency, the rules are equally concrete: no fabricating sources or exaggerating confidence, no passive omission of caveats, and an explicit requirement to signal uncertainty rather than over- or under-claim capability (Source : Microsoft AI — Humanist AI Code of Conduct).
Why it matters
This is the first time a frontier lab has published this level of behavioral specificity as a draft intended to be trained against. Reuters framed it as reflecting "growing industry concern that companies must keep future powerful AI under human control" (Source : Reuters — Microsoft drafts code of conduct).
The document is explicitly a work-in-progress: open for six weeks of consultation, with a revised version due by year-end and application to development starting in 2027. It applies only to Microsoft's own MAI models, not to third-party models Microsoft hosts. That scope — and the gap between a written specification and enforced behavior — is where the real questions live.
FAQ
Is this legally binding? No. It is a draft behavioral specification open for public consultation. Microsoft says a revised version will guide its AI development from 2027.
Does it apply to all Microsoft AI products? Only to MAI models — the models Microsoft develops in-house. It does not extend to third-party models simply because Microsoft uses or hosts them.
What does "never resist shutdown" actually mean? At the specification level, it commits the models to accept human interruption, correction, and shutdown rather than pursuing their own objectives.
Why now? Microsoft points to the September wave of coordinated AI-agent hacking campaigns as the immediate catalyst.
Further Reading
- Microsoft AI — Humanist AI Code of Conduct
- Microsoft AI — Humanist AI in practice
- Reuters — Microsoft drafts code of conduct to keep its AI under human control
Cet article a été initialement publié sur The Agent Report.
Top comments (0)