Cognitive surrender, defined
/ Bridgekeeper team
In a forthcoming paper, Steven Shaw and Gideon Nave at Wharton give a name to a behavior most engineering teams have already seen: cognitive surrender.
Cognitive surrender is the uncritical adoption of an AI’s answer in place of one’s own reasoning. It is not the same as delegation. Delegation is “you handle this, I will check it.” Surrender is “you handle this.” The check never happens.
Shaw and Nave ran three studies with 1,372 participants and 9,593 trials. Participants answered reasoning problems drawn from the Cognitive Reflection Test. Half had access to an AI chatbot. The researchers manipulated the chatbot: half the time it gave the correct answer, and half the time it gave a deliberately wrong one.
The results are stark.
- When the AI was correct, participant accuracy jumped to 71%, up from a 45.8% baseline.
- When the AI was wrong, accuracy fell to 31.5%, well below baseline. Participants followed the faulty AI 79.8% of the time.
- Confidence rose either way. Participants reported being 11.7 percentage points more confident in their AI-assisted answers, including the wrong ones.
The authors propose extending Kahneman’s dual-process model with a third system. System 1 is fast and intuitive. System 2 is slow and deliberate. System 3, they argue, is AI consultation, and it is becoming a default that bypasses System 2 entirely rather than a tool that System 2 reaches for.
For software teams the implication is uncomfortable. Code review is supposed to be a System 2 activity: deliberate, effortful, checking the model’s work. If the reviewer surrenders to the AI’s framing of what the diff does, the review is theater. The thumbs-up is real. The understanding is not.
Bridgekeeper is built to make surrender harder. We ask the reviewer to predict what a salient change does before we show them the AI’s explanation. That single ordering, predict first and reveal second, turns a System 3 autopilot moment back into a System 2 check. You cannot bluff a prediction the way you can bluff a thumbs-up.