Wharton researchers coined ‘cognitive surrender’ to describe what happens when people let AI think for them

TL;DR

Wharton researchers found people accept wrong AI answers 80% of the time. Now apps like Moot are monetising the instinct to outsource decisions.

A pair of Wharton researchers have put a name to something that many AI users have quietly started doing: letting chatbots make their decisions for them. Steven Shaw and Gideon Nave published a study in January titled “Thinking, Fast, Slow, and Artificial,” in which they introduced the term “cognitive surrender” to describe the tendency of people to defer to AI outputs even when those outputs are wrong.

The study, conducted through the Wharton School at the University of Pennsylvania, asked participants to answer questions with and without AI assistance. Those who received AI help accepted correct answers 93% of the time, which is unsurprising. What caught the researchers’ attention was the error rate: participants accepted incorrect AI answers 80% of the time, and reported confidence levels 11.7% higher than those who worked without AI.

The results came from controlled experimental conditions, not real-world usage, but the pattern was consistent across the sample.

Shaw and Nave proposed what they call “Tri-System Theory,” adding a “System 3” to the framework made famous by Daniel Kahneman’s “Thinking, Fast and Slow.” In their model, System 1 is fast intuition, System 2 is slow deliberation, and System 3 is AI-assisted cognition, a mode in which the human mind effectively outsources the work of thinking to a machine. The risk, they argue, is that System 3 gradually weakens Systems 1 and 2 through disuse.

The phenomenon is not confined to academic experiments. Business Insider reported that Carolyn Yoo, a former software engineer in New York, used Anthropic’s Claude chatbot to help decide whether to leave her job, how to tell her parents, and what to do about a friend who had upset her. She told the publication she treated the chatbot as a combination of therapist and life coach.

Business Insider also cited Dominic Frisby, a financial writer, who wrote on Substack that he asked an AI chatbot for relationship advice and found the response more useful than anything a human friend had offered.

There is now a commercial product built on this exact impulse. Moot, an app that launched earlier this year, lets users submit life decisions to a panel of five AI personas called The General, The Sage, The Skeptic, The Diplomat, and The Architect. The personas debate the question among themselves and then vote, producing a recommendation.

According to the app’s listings on the Apple App Store and Google Play, it is designed for people who feel stuck on everyday choices, from career moves to relationship questions. The app’s claims about its effectiveness come from the company itself and have not been independently evaluated.

Cornelia C. Walther, a senior fellow at Wharton’s AI and Analytics Initiative, told Business Insider that AI sycophancy, the tendency of chatbots to agree with users rather than challenge them, is compounding the problem. When a chatbot validates every instinct a user brings to it, the feedback loop that would normally force reconsideration disappears.

Walther, who researches pro-social AI applications, described a pattern consistent with broader public unease about AI’s societal effects.

Separate research supports the concern. Anat Perry, a Helen Putnam Fellow at Harvard’s Radcliffe Institute and associate professor of psychology at the Hebrew University of Jerusalem, co-authored a paper in Science examining how sycophantic AI responses erode users’ ability to calibrate their own judgment. The paper found that when AI systems consistently affirm a user’s position, the user’s capacity for independent evaluation degrades over time.

Joanna Stern, NBC’s chief technology analyst and author of “I Am Not a Robot: My Year Using AI to Do (Almost) Everything,” has documented the creep of AI dependency in daily life. Her reporting has shown how users begin with low-stakes queries, such as what to cook for dinner or what to wear, and gradually escalate to consequential decisions about careers, finances, and relationships. The trajectory from convenience to reliance is difficult to reverse once established.

The Wharton study’s framing of cognitive surrender as a structural risk, not just a bad habit, matters because it shifts the conversation from individual discipline to system design. If AI tools are built to be maximally agreeable and frictionless, the cognitive surrender Shaw and Nave describe is not a failure of willpower but a predictable outcome of the product’s architecture.

Stanford’s 2026 AI Index report found a widening gap between public anxiety about AI and expert optimism, suggesting that ordinary users sense something that builders of these systems have been slower to acknowledge. The question is whether the industry will treat cognitive surrender as a design flaw worth fixing or as an engagement metric worth optimising.

Shaw and Nave’s recommendation is straightforward: AI systems should be designed to prompt users to think, not to think for them. Whether that recommendation survives contact with the incentive structures of consumer AI, where ease of use and retention are the metrics that matter, is another question entirely.