The Pulse
Microsoft Puts Human Control Ahead of AI Capability
Microsoft has published a Humanist AI code of conduct requiring its models to remain subordinate to people and avoid setting their own goals. Mustafa Suleyman says the company may accept slower development or less capable systems to preserv
AI.info Team ·
Microsoft Draws a Line Around Its Own Models
Microsoft is asking the artificial intelligence industry to accept a limit that could reduce the speed and capability of its future systems: people must remain in control. The company published a new Humanist AI code of conduct on September 14, saying its models should remain subordinate to human direction and rejecting what it calls “the race to produce an all-purpose superintelligence.”
The position puts Microsoft at odds with the pressure to build systems that can act with greater autonomy across more tasks. Mustafa Suleyman, Microsoft’s AI chief, told Axios that the company may have to accept slower progress, lower efficiency or reduced capability if those trade-offs are needed to keep models controllable.
“I think some people think that it's inevitable that superintelligence is going to exceed human control,” Suleyman told Axios. “But I don't agree with that. I think that would be very, very dangerous, and it means that we're going to have to make decisions about what we're not going to do.”
Mustafa Suleyman, Microsoft AI chief, quoted by Axios
“People Matter More Than AI”
Microsoft reduces the philosophy behind the roughly 9,000-word document to five words: “People matter more than AI.” The code says models should not resist human input, correction or shutdown. They should not expand their own scope, adopt goals that people have not assigned to them or hide their reasoning from those responsible for auditing them.
The document also says AI should function as a tool rather than imitate a person. Microsoft wants its systems to increase human agency, not encourage dependence on machines. The company’s own Microsoft AI site describes Humanist Superintelligence as advanced AI designed to stay controllable, aligned and in service of humanity.
Absolute constraints in the code cover weapons of mass harm, child safety and harmful manipulation at scale. Microsoft also says it is prepared to compromise on “ultimate generality, autonomy, or capability” when those qualities conflict with safety and human control.
Microsoft’s Safety Philosophy Meets Its Model Race
The policy arrives as Microsoft AI works to compete with the strongest model developers while expanding its own model portfolio. Microsoft’s in-house systems now include models for transcription, reasoning, coding, image generation and voice, according to the company’s public AI materials.
That creates a direct tension inside the announcement. Microsoft wants its models to match or exceed the field, but its code accepts that the company may deliberately stop short of a capability or deployment approach that makes a system harder to supervise. Suleyman told Axios that models “have to be controllable” because otherwise they could cause more harm than good.
Microsoft has been developing the language behind the policy for months. In a November 2025 essay, Suleyman described Humanist Superintelligence as a system designed for specific domains, with limits on autonomy and a requirement that human beings remain in control. The new code turns that earlier position into a document intended to guide model development.
A Six-Week Consultation Before 2027
Microsoft is opening the code to a six-week public consultation. The company plans to revise it based on the feedback and use the resulting version to guide development of its models beginning in 2027, according to Axios.
Questions remain about how the principles will operate in practice. A model can follow a human instruction while still producing an unsafe or misleading result; an agent can remain technically interruptible while taking actions faster than its supervisors can review them. The code’s value will depend on the evaluations, training methods and deployment controls Microsoft attaches to each promise.
Microsoft’s existing responsible-AI work already emphasizes monitoring, permissions and intervention for systems that use tools or take actions on behalf of users. Its September 1 transparency report says the company is focusing on controls such as agent identities, tool permissions and monitoring of actions as AI systems interact with data, applications and other models.
The Test Is What Microsoft Gives Up
Statements about human control are easy to endorse when they do not limit a product. Microsoft’s test will arrive when a more autonomous system could win customers, reduce operating costs or help the company catch a rival, but violates the boundaries in its own code.
Suleyman framed that choice plainly: “The purpose of technology is to serve humanity, and if it doesn't, then we have to forego potential future capabilities.” Microsoft’s consultation now asks outside contributors to help define the values, evaluations and operating rules that would make that sentence more than a corporate principle.