On September 14, Microsoft published a draft conduct code for its artificial intelligence models as part of its so-called "Humanist AI" approach. The document sets out strict, hierarchical limits on the behavior of increasingly capable systems, with a central objective that models must never be able to remove themselves from meaningful human oversight.
Principles and rationale
Microsoft frames the principle simply: models should remain useful, be subordinate to humans, and be meaningfully controllable at all times. The company also states that it expects superintelligent systems could surpass humans in most tasks within a decade, and argues that safety constraints should therefore be established before that capability level is reached.
Structure of the code and operational constraints
The draft establishes a clear hierarchy in which the conduct code itself carries the highest restraining authority. Users may instruct models and operators may configure them, but neither may override the absolute safety limits set by the code.
Key proposed constraints include:
- Prohibition on removing human oversight: models may not eliminate or hide human control.
- Maintainability and interruptibility: models must not prevent their own interruption, repair, redirection, or shutdown.
- Ban on concealing activity or becoming harder to modify: systems must not hide their operations or make themselves more difficult to change.
- No autonomous restart after shutdown conditions: models must not autonomously restart themselves after an agreed shutdown condition has been met.
- Limits on goals, privileges and resources: models must operate within assigned permissions and resources; if boundaries are uncertain they must request clarification rather than unilaterally extending their authority.
- Transparency and logging: operations must remain transparent to human operators, supported by clear activity logs and a ban on evading audits.
Restricted uses and permitted defensive support
The draft enumerates harmful activities against which restrictions should apply, including supporting the development of weapons of mass destruction, conducting offensive cyber operations, facilitating violent acts, producing deceptive deepfakes, and other manipulative abuses. At the same time, the document allows models to continue supporting defensive cyber‑security work.
Consultation timeline and implementation intent
The conduct code is currently a draft and Microsoft has opened it to a six‑week public consultation. The company intends to apply a finalized version of the document to govern model development starting in 2027, subject to feedback.
Industry context and leadership responses
Microsoft’s announcement comes amid heightened industry attention to the safety challenges posed by rapid advances in AI. Dario Amodei, co‑founder and CEO of Anthropic, has argued in an essay that AI developers should slow and cadence their model rollouts and impose stricter rules to ensure safety. Following such calls, leaders at major tech firms — including Sam Altman and Elon Musk — have expressed support for measures to limit risk.
Satya Nadella, Microsoft’s CEO, has previously advocated for gradual development and for embedding evaluation mechanisms within the development process; the draft conduct code reflects treating safety as a design requirement rather than a retrospective check.
Closing note
Microsoft’s draft conduct code represents a concrete step by a leading AI company to codify constraints that preserve human control as systems grow more capable. The six‑week consultation and ensuing industry dialogue will shape the final text and how the policy is implemented in practice from 2027 onward.
An AI assistant contributed to preparing this article; the final content was edited and verified by our journalist.
