Nadella Proposes AI Safety Controls for Reviewing and Stopping Tasks

Human oversight while AI works

Microsoft CEO Satya Nadella proposed that people be able to review artificial intelligence actions and stop a model mid-task. According to TechCrunch, he outlined his AI safety views in an October 10, 2026, post on X.

Nadella argued against treating advanced AI as a system people cannot see into, whose recommendations and actions they simply accept or reject. His proposed alternative gives human oversight a role during the work itself.

Records explain actions; intervention can stop them

According to TechCrunch, Nadella called for every meaningful model action to produce tamper-proof evidence that humans can read. In plain terms, people would have an action record designed to resist alteration, rather than just a recommendation to assess.

He also proposed that an authorized person always be able to pause or shut down a model during a task. That serves a different purpose: records support review, while intervention lets someone interrupt work before it finishes.

Containment starts outside the model

Nadella proposed separating the AI model from its harness—the surrounding system that organizes its work—and placing safeguards outside the model, TechCrunch reported. Under that approach, controls would be a separate part of the system's design.

He advocated assuming a model is compromised and containing it from the outset. This is a proposed design rule for limiting the model, not a finding that every model has been compromised.

These are proposed requirements. The supplied reporting does not establish how the protections would work in practice or whether Microsoft plans to adopt them.