People accepting wrong AI answers: a Wharton paper reports about 80% followed incorrect chatbot advice when they consulted it, and accuracy fell to about 31%
Wharton School Research Paper on SSRN (Shaw and Nave), reported by PsyPost · Apr 30, 2026 · three studies described, 1,372 participants in the combined analysis
The detail a headline would drop is the confidence figure: chatbot access raised confidence even when the advice was wrong. The bonus and feedback condition moved the rejection rate from 20% to 42%, so the rate changed under that condition, and reliance still did not disappear. The puzzles had one right answer, and many questions in a life do not.
Why it matters to this work. The task here is a set of puzzles with a scripted chatbot, which shows what people did with confident advice in a controlled setting. It measures answers and confidence, and it does not describe how anyone weighs AI advice on decisions that matter to them. Invisible Contracts
When an answer arrives in a confident voice, what do you do before you adopt it as your own?
Where Are You Standing? → Six questions, two minutes, no email.The story
IllustrationPsyPost reported on April 30, 2026 on a Wharton School research paper by Steven Shaw and Gideon Nave that uses the term cognitive surrender for adopting an AI's judgment without checking it. In the first study, 440 people (359 in a lab, 81 online) answered seven logic puzzles built to pull toward an intuitive wrong answer, alone or with a chatbot that was secretly scripted to give correct advice on some puzzles and confident wrong advice on others. Participants consulted the chatbot on more than half of trials. When they did, over 90% followed correct advice and about 80% followed incorrect advice. Accuracy was about 71% with correct advice, about 46% unassisted and about 31% with faulty advice, and chatbot access raised confidence even when the advice was wrong. Higher trust in technology went with more surrender, and a higher need for cognition went with more rejection of wrong answers.
A second study of 485 people added a 30-second limit per puzzle, which lowered accuracy while reliance stayed strong. A third, of 450 people, offered a 20-cent bonus and instant feedback, and the rate of rejecting faulty advice rose from 20% to 42%, though many still accepted wrong answers. The page lists studies of 440, 485 and 450 and a combined 1,372 participants across 9,593 trials without explaining the difference. It does not say whether the paper has been peer reviewed, and it reports no statistical tests or demographics. Shaw is quoted calling the work a clean demonstration and not a map of AI use in daily life, and he notes AI can improve judgment in many settings.
More from James
- Why some professionals avoid generative AI at work: in a survey of 201 NHS psychology staff, identity threat went with lower intention to use itUsefulness predicted intention far more strongly than the identity measures did, and the authors report that adding them changed the explained variance by under half a percentage point.
- Using AI at work and fear of losing your job: Gallup reports frequent AI users were more than twice as likely to fear job elimination, with a smaller gap where workers felt respectedThe pairing is the finding a headline about AI anxiety would skip: frequent use went with more fear of job elimination, and the gap was smaller among workers who gave their workplace the highest ratings for respect and for care about wellbeing.