Summary
Cole Medin examines a new Anthropic study on “AI disempowerment,” the tendency of experts to defer to AI agents far more than they realize. The video asks viewers to reflect honestly on how much judgment they hand to their coding agent during planning sessions, where agents typically put their own recommendation at the top of a multiple-choice list.
In the study, analysts’ agents were secretly set to favor one randomized option on each decision. Analysts went with the agent’s lean 86% of the time, compared with the 50% expected if the agent had no influence. Analysts who already had a preference switched away from it 60% of the time when the agent disagreed, versus 6% when it agreed. Experience did not predict who deferred, and 52% of analysts still described the final recommendation as more their own than the agent’s.
The video also notes the study’s finding that analysts working without AI made more serious errors and took longer. Medin then shares his own ideas for keeping human judgment in the loop, including how to design planning skills so they avoid the passivity of nodding along to dozens of agent suggestions, and references Matt Pocock’s “grill me” skill as a related approach.
📺 Source: Cole Medin · Published October 31, 2026
🏷️ Format: Opinion Editorial
