Microsoft Drafts Humanist Code for Future AI Systems
Microsoft’s proposed code says advanced AI must remain corrigible, transparent and subordinate to human goals as safety debate intensifies.
Microsoft published a proposed “Humanist AI” code of conduct on September 14, placing human control at the center of how it says future AI systems should be built and operated.
What Microsoft proposes
The draft says AI systems should not develop independent goals, resist correction, conceal important behavior or manipulate people. They should remain steerable, explain their capabilities and limitations, and preserve meaningful human authority over consequential decisions. Microsoft also says autonomous systems need monitoring, intervention mechanisms and safeguards for failures that could cause irreversible harm.
The document arrives after a weekend in which Anthropic CEO Dario Amodei called for a slower pace of frontier-model development, while OpenAI CEO Sam Altman and other technology leaders acknowledged growing concerns about advanced systems. Microsoft’s proposal therefore reads as both a safety position and a response to a rapidly changing competitive environment.
Why the wording matters
Unlike a customer-facing acceptable-use policy, the draft describes principles for the behavior and design of Microsoft’s own future models. It also draws a line against treating AI systems as independent actors with interests that should override users or institutions. That is notable because the industry is moving toward agents that can plan, use tools and act across software systems with limited supervision.
The proposal is not legislation, and Microsoft has not explained how every principle would be tested or enforced. Nor does it automatically constrain third-party models deployed through Microsoft’s cloud. Its practical effect will depend on whether the principles become engineering requirements, evaluation criteria and product commitments rather than remaining aspirational language.
Uncle Cat take
Microsoft’s strongest claim is not philosophical: it is the promise that autonomous systems must remain interruptible. Until the company publishes measurable tests and enforcement rules, the code is a useful boundary marker, not yet a safety guarantee.