The Microsoft AI division sets limits that no operator can loosen — CBRNE weapons, cyber attacks, mass manipulation — and opens its text to six weeks of public comment.
On September 14, 2026, Microsoft AI published a code of conduct to govern its in-house models. The document lists absolute prohibitions, presented as non-deactivable, including by companies that deploy these models: assistance in the development of chemical, biological, radiological, nuclear or explosive weapons, production of child sexual abuse content, support for offensive cyberattacks, mass manipulation and disinformation campaigns, and any action aimed at escaping human control.
The text goes beyond the list of prohibitions by establishing a doctrine: models must remain subordinate to human decisions, amplify users' capacities rather than replace them, respect their autonomy over important choices, remain transparent about their artificial nature and avoid presenting themselves as substitutes for human relations. Microsoft explicitly rejects the idea that its models can claim a form of consciousness.
The document is published in a consultation version, open for comments for six weeks before revision. It commits the institution — no individual signature appears — and is part of a sequence in which the laboratories each publish their own framework upstream of a regulation that is still incomplete.
Read the full article on the original source
Go to the source →