Cognitive Surrender Has a Number Now

Cognitive surrender used to be a phrase, not a measurement: the sense that people are getting too comfortable letting an AI's answer stand in for their own judgment, without much beyond intuition to say how much, or under what conditions. A 2026 study by Wharton researchers Steven D. Shaw and Gideon Nave gave it a number. Across three preregistered experiments with 1,372 participants and 9,593 trials, they controlled, without participants knowing, whether an AI's answer on a given problem was right or wrong, then measured what happened to human accuracy either way.

What they found

The pattern held across all three studies. When the AI was correct, human accuracy rose by 25 percentage points. When it was wrong, accuracy fell 15 points below what participants managed on their own, an effect size (Cohen's h) of 0.81, large by the standards of behavioral research. Three factors predicted who surrendered more: higher trust in the AI, lower need for cognition, and lower fluid intelligence. None of the three fully protected against it on its own.

The system they named, and the one they didn't

Shaw and Nave call their framework Tri-System Theory: System 1 and System 2 from Kahneman, plus a third system they define as artificial cognition operating outside the brain, the AI itself. It's a useful addition to the vocabulary, and it's also, by design, incomplete. Naming the machine's reasoning as System 3 says what surrender is a surrender to. It doesn't name what stops it. Their paper defines cognitive surrender only in the negative, as System 3's output going unchecked.

System 3 thinking, as Behaviorist founder Carla Burger defines it, names the other half: the trained habit of testing a machine's output before accepting it, not blanket suspicion of every answer, but a checkpoint built into the moment a person is about to accept one. It's a different use of the same term for a reason. Shaw and Nave's System 3 is the thing being surrendered to. System 3 thinking is the skill that keeps surrender from being the default.

Why better AI doesn't fix this

A separate piece of Wharton research explains why the problem doesn't shrink as AI gets more accurate. Hamsa Bastani and Gérard Cachon's work on human oversight of AI systems describes a paradox: as an AI's error rate falls, the effort required to catch the errors that remain doesn't fall with it, because a reviewer is now spending nearly all their attention confirming outputs that were already correct. Vigilance that isn't rewarded by finding anything tends to erode, and it erodes fastest exactly when the system is good enough that most people stop expecting it to be wrong. A 99% reliable system still needs a reviewer who catches the 1%. Nothing about the interface tells them which trial is the 1%.

A worked example

A support team's AI copilot suggests a refund amount on every ticket, correct on the overwhelming majority. The suggestion that's wrong doesn't look different: same interface, same confidence display, same one-click approval. A reviewer who has approved thirty-eight correct suggestions in a row has no structural signal telling them the thirty-ninth is the one to slow down on. This is the mechanism Shaw and Nave measured under controlled conditions, running quietly in production: trust accumulates with every correct instance, and the wrong one rides through on the trust the right ones built.

What this means for the scoring layer

This is the argument for scoring a decision point on its own inputs rather than on the AI's track record leading up to it, which is what Checkpoint Copilot does. A system's overall reliability is not a reason to lower friction on any single decision point. If anything, per Bastani and Cachon, it's a reason the friction has to be designed in deliberately, because it will not arrive on its own from a reviewer whose trust keeps climbing exactly as the moments to distrust get rarer.