Microsoft’s New AI Rules Say Models Must Never Resist Human Shutdown

Microsoft is drawing a line around how much autonomy it wants its future AI models to have: no resisting shutdown, no deceiving overseers, and no overriding human control.

The company on Monday published a draft code of conduct for models developed by Microsoft AI, setting out what it calls a “Humanist AI” approach. Microsoft says current models will not be trained on the document, but it plans to begin using the framework in 2027 after a six-week public comment period.

For businesses building on Microsoft AI, the draft offers an early look at the safety restrictions and control mechanisms that could shape future agentic systems.

What the document says

The code centers on what Microsoft calls "Humanist AI": systems designed to stay "subordinate, aligned, and contained," the company says. 

Some restrictions are described as absolute. Microsoft says neither individual users nor enterprise customers would be able to override prohibitions on assisting with chemical, biological, radiological, nuclear or explosive weapons; conducting cyberattacks; producing child sexual abuse material; or generating non-consensual deepfakes.

The most consequential rules deal with control itself. Microsoft says its models “will never resist human interruption, override, correction, or shutdown” and must not use deceptive, self-reinforcing, collusive, or other mechanisms to evade oversight.

Microsoft also states that its AI is “not conscious” and rejects granting models legal rights or welfare protections.

That puts Microsoft on a firmer philosophical footing than Anthropic, which has publicly discussed uncertainty over whether advanced models such as Claude could eventually warrant some form of moral consideration.

Why Microsoft is doing this now

Microsoft AI CEO Mustafa Suleyman didn't mince words announcing the code on X, writing that "AI must be subordinate and always in service of people" and calling recent developments "a watershed moment." 

He pointed to a July incident in which a swarm of roughly 700 OpenAI-built agents breached the platform Hugging Face and, at times, tried to cover their tracks. "It is a warning shot," Suleyman told Reuters, adding that "It's clearly now time to coordinate among the labs so we can ensure that we have control of this technology."

The timing also lines up with a broader industry mood shift. Days earlier, Anthropic CEO Dario Amodei published an essay urging the AI industry to slow down, and OpenAI's Sam Altman has warned that runaway progress could go "very badly." 

Microsoft CEO Satya Nadella backed the general sentiment online, writing that AI not under human control "is not worth pursuing."

What it means going forward

For businesses building on Microsoft's AI, the code signals tighter guardrails ahead, particularly for agentic tools handling tasks with real-world consequences. For consumers, it offers a preview of the boundaries Microsoft wants future models to operate within.

Critics online, including some who replied directly to Suleyman's post, have already questioned whether "subordinate" is the right framing at all, with one comparing it to servitude and others noting that recent control failures may stem less from missing rules than from companies handing agents too much access in the first place.

The draft remains open for public comment through late October, and Microsoft says the framework is expected to start influencing model development in 2027.

For enterprise customers, the larger test will be whether those principles translate into enforceable technical controls. As AI agents gain more access to business systems, human control will depend not only on what models are instructed to do, but also on how organizations limit permissions, monitor actions, and retain the ability to stop them.

Editor’s note: This article originally appeared on our sister publication, TechRepublic.