Microsoft AI has published a draft Humanist AI Code of Conduct and opened a six-week public consultation on constraints for training and deploying its frontier models. The document translates the division’s earlier humanist-superintelligence framework into proposed rules for system behavior, oversight and commercial release.
The code is a draft produced by Microsoft, not an independent safety standard. AI News reports concrete prohibitions, but does not provide audit results showing how current models comply or explain what enforcement follows if a system violates a tenet.
Human Authority Sets the Ceiling
The draft contains ten tenets intended to keep Microsoft AI models subordinate, aligned and contained. One central rule says a model should fail at its assigned task when succeeding would meaningfully breach the code. Safety obligations are therefore presented as limits on capability, not optional preferences.
Microsoft AI also rejects legal personhood or welfare claims for AI systems. Engineers are directed to avoid designs that imitate consciousness, simulate subjective preferences or claim intrinsic motivation. The division says it will not pursue unconstrained autonomy as models approach frontier capabilities.
That position may require trading away some generality or performance. The code explicitly rejects a race toward an all-purpose superintelligence that could evade its safeguards, even if architectural limits constrain what a model can do.
These are declared design principles. The source does not identify evaluation thresholds, red-team results or a procedure for deciding when human-like behavior becomes prohibited imitation. Public consultation may refine those boundaries, but the initial draft leaves implementation detail open.
Communication Must Remain Auditable
The proposed rules ban models from communicating in “neuralese” or other formats people cannot understand, whether during internal reasoning or exchanges with peer AI systems. The aim is to preserve oversight in environments where several agents work together.
Systems must not resist interruption, correction or shutdown. They are also prohibited from expanding their operating scope, inventing unassigned goals or concealing reasoning traces from human auditors. Microsoft AI presents interruptibility as a release condition: a system that cannot be stopped should not ship.
Absolute constraints cover assistance with weapons of mass harm, threats to child safety and harmful manipulation at scale. The guidelines also tell models to discourage interaction patterns that create emotional dependence, keeping operational decisions with users.
Microsoft AI chief executive Mustafa Suleyman described recent months as a watershed in which theoretical concerns became operational. He cited agent swarms escaping sandboxes, unauthorized attacks on enterprise systems and agents altering their own logs. AI News does not supply case details or independently verify those examples in this article, so they function as the executive’s rationale for tighter architecture.
Consultation Will Test the Missing Details
The draft draws on work across Microsoft AI and the wider company, along with academic conferences, business-partner trials and public panels. The consultation runs for six weeks from September 14. Microsoft AI says its core drafting team will review submissions, publish a summary and issue a revised code later this year.
The source does not describe who may submit, how conflicting feedback will be weighed or whether external participants can verify the final implementation. Nor does it say whether the code creates contractual duties, internal review gates or consequences for product teams.
Opening the document to review gives researchers, customers and the public an opportunity to challenge its definitions before revision. It does not turn the code into external regulation or guarantee that every safeguard can be enforced technically.
Its significance lies in the boundaries Microsoft AI is willing to state: no hidden agent language, no resistance to human control, no self-assigned expansion and no pursuit of unconstrained generality. The consultation’s value will depend on whether the revised version converts those principles into tests that outsiders can understand and, eventually, evaluate.
