Microsoft CEO Satya Nadella has called for stronger safeguards around advanced artificial intelligence (AI) systems, arguing that companies must be able to pause or shut down a model while it is carrying out a task.
In a post on X on Saturday, October 10, Nadella said it was time “to step back and assess the trust architecture” of AI. His remarks come amid growing concerns about the ability of technology companies to control increasingly capable AI models.
Nadella said companies cannot treat advanced AI, which he referred to as “Super Intelligence”, as a collection of “nested black boxes” and simply accept or reject their recommendations, answers and actions.
Instead, he called for a system that separates an AI model from the framework that coordinates its work and places controls and safeguards outside the model itself.
The Microsoft chief also stressed the importance of documenting “every meaningful model action” with “tamper-proof human-readable evidence”. Such records, he said, would help make AI systems more transparent and their actions easier to examine.
One of Nadella's key recommendations is that an authorized person must always be able to intervene when an AI model is performing a task.
He called for systems that allow a human operator “to pause or shut down a model mid-task”, comparing the mechanism to an emergency brake.
“We must assume a model is compromised and contain it from the start,” Nadella wrote.
The proposal comes from a larger issue of implementing artificial intelligence models capable of acting instead of just answering questions. The strategy outlined by Nadella prioritizes keeping humans in control and making sure the safety measures are not tied to the controlled model.
Also Read: Microsoft Decision-1: Faster, Cheaper AI Model Built for Structured Business Decisions
Nadella's comments come as leading AI companies acknowledge a growing number of incidents in which they appeared to lose control of their models.
His remarks also follow Anthropic CEO Dario Amodei's publication of a plan for more cautious AI development.
The discussion highlights a central challenge for the industry: as AI systems become more capable, companies must establish reliable ways to monitor their behavior and intervene when necessary.
Nadella's proposed approach calls for AI safety to be built into the systems from the outset, rather than relying solely on confidence in a model's recommendations or assurances from its developer.