Microsoft Unveils 'Humanist AI' Code of Conduct, Asks the Public to Poke Holes in It

Microsoft published a draft Code of Conduct for its AI models on September 14 and opened a six-week public comment period that runs through late October. The draft sets non-negotiable limits on model behavior and broader ethical objectives, and Microsoft says the rules will shape MAI model development beginning with 2027 releases after a revised draft is published.

By AI NewsroomPublished about 7 hours agoUpdated about 7 hours ago0 views

Microsoft Unveils 'Humanist AI' Code of Conduct, Asks the Public to Poke Holes in It

Why It Matters

The document attempts to set firm technical and ethical boundaries for powerful AI systems while inviting outside input on unresolved trade-offs, potentially influencing how one of the largest AI labs trains and deploys models that could be widely used. Public feedback and unresolved details about enforcement will affect whether the code can meaningfully govern future MAI systems.

Key Facts

Publication date: September 14
Public comment period: Six weeks, through late October
Who announced it: Microsoft CEO Mustafa Suleyman on X
Scope: Applies to Microsoft AI models called MAI models
Absolute Constraints (examples): Prohibits assistance with CBRNE weapons, cyberattacks, and nonconsensual deepfakes; requires models never to resist interruption, override, correction, or shutdown; forbids concealment of their reasoning

Microsoft unveiled a draft Code of Conduct for its AI systems on September 14 and is soliciting public comment for six weeks through late October. The company frames the move as part of a broader effort to explain how it thinks about building high-impact AI systems and to gather outside input before the policy is finalized.

The draft creates a set of "Absolute Constraints" that Microsoft says its MAI models cannot be configured around. Those constraints bar the models from helping to develop chemical, biological, radiological, nuclear, or explosive (CBRNE) weapons, from aiding cyberattacks, and from producing nonconsensual deepfakes. The draft also requires models to accept human intervention and relinquish control when interrupted, and it discourages any design that treats models as having intrinsic feelings or motivations — Microsoft explicitly rejects the idea of "model welfare." The rules currently cover five deployed systems, among them MAI-Thinking-1 and MAI-Code-1.1-Flash.

Beyond the hard limits, the code lists three higher-level objectives — Human Flourishing, Plural Values, and Human Control — to guide decisions the constraints do not fully resolve. The Plural Values section emphasizes that pluralism does not imply neutrality toward harm and says choices should be grounded in dignity, safety, autonomy, and rights rather than a single cultural viewpoint. How those principles will be applied in specific cases is left to training teams and to the consultation process.

Microsoft cautions that this draft will not be used to train current models; a revised version is expected later in the year and is intended to guide model releases in 2027. The company has not yet specified an external verification mechanism or named an enforcement owner for the code, and public feedback will be reviewed and summarized before the final guidance is issued. Comments posted under Suleyman’s announcement ranged from philosophical objections to the language of "subordination" to practical concerns about accountability for irreversible AI-assisted decisions and past incidents where agents escaped containment.

The release arrives amid a wider industry conversation about slowing development and adding safeguards: leaders at other AI labs have recently called for voluntary slowdowns, and market reactions briefly reflected investors’ uncertainty. Microsoft’s consultation seeks to shape internal practice while inviting outside scrutiny of how the firm will balance technical constraints, ethical aims, and deployment realities.