Wharton (2026) — AI, reasoning, and cognitive surrender

Summary:

The paper studies “cognitive surrender”: when people let AI do the thinking and accept its answers with little checking — even when they’re wrong.

Shaw and Nave (Wharton) propose “Tri‑System Theory,” adding AI as a third thinking system alongside fast intuition and slow deliberation. In three preregistered experiments, 1,372 participants answered seven Cognitive Reflection Test trick questions (9,593 total trials), either alone or with an embedded ChatGPT assistant.

Crucially, the researchers controlled the assistant to give either the correct answer or a confident wrong one. They tracked AI use, whether people followed it, accuracy, and confidence; later studies added time pressure and then incentives plus immediate feedback.

Results: participants used AI on over half of trials and usually copied it — following correct AI ~93% of the time and wrong AI ~80%. Accuracy rose when AI was right and dropped below the no‑AI baseline when AI was wrong. Time pressure hurt baseline performance; good AI buffered it, bad AI still harmed it. Incentives and feedback reduced (but didn’t eliminate) reliance on bad AI and improved accuracy overall.

AI also boosted confidence even when wrong.

Higher AI trust increased copying; higher need for cognition and fluid intelligence helped people resist faulty AI.

Limits: these are lab CRT puzzles, not real‑world decisions, and “AI correctness” was experimentally forced. The paper is well‑disclosed and preregistered, with clear, visualizable gaps between no‑AI, accurate‑AI, and faulty‑AI conditions.