Cognitive surrender

Cognitive surrender is the tendency to take an answer produced by an artificial intelligence (AI) system as one's own, with little or no scrutiny, instead of reasoning through the problem oneself. The term was introduced in January 2026 by Steven D. Shaw and Gideon Nave of the Wharton School of the University of Pennsylvania, who defined it as the tendency to defer "judgment, effort, and responsibility" to an AI system's output, particularly when that output is delivered fluently, confidently, or with minimal friction.^[1:1] They contrast it with cognitive offloading, in which a person hands a discrete task to a tool, such as a calculator, while their own reasoning remains in charge. In cognitive surrender, they argue, the person stops evaluating and simply adopts the machine's judgment.^[1:2]

Shaw and Nave place the concept within what they call Tri-System Theory, which extends the familiar dual-process picture of fast intuition ("System 1") and slow deliberation ("System 2") with a third, external "System 3": the artificial cognition supplied by AI tools.^[1:3] In three preregistered experiments with 1,372 participants, people solving reasoning puzzles could consult a chatbot that had been secretly set to give either the right answer or a plausible wrong one.^[1:4] Participants who consulted it followed its advice on about four out of five trials even when it was wrong, and those with AI access reported higher confidence than those without it.^[1:5][1:6]

The term attracted press coverage in early 2026, and other researchers soon adopted it: to interpret declines in student learning, to describe AI users' reduced willingness to say "I don't know", and to frame the challenge of keeping humans meaningfully in charge of AI agents.^{[2][3:1][4:1][5:1]} Its originators stress that surrender is not always irrational, because deferring to a more accurate system can be the best choice.^[1:7] The concern is that people do this without knowing whether the system is right.

Origin and definition

Shaw and Nave introduced the term in a paper titled Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender, posted to the PsyArXiv preprint server on January 12, 2026 and also circulated as an SSRN working paper.^[1] The title echoes Daniel Kahneman's book Thinking, Fast and Slow, which popularised the System 1 / System 2 vocabulary.^[1][2]

The authors define cognitive surrender as "the behavioral and motivational tendency to defer judgment, effort, and responsibility" to an AI system's output. In their account, the effect is strongest when the answer arrives fluently and confidently, and when accepting it takes little effort.^[1:1] They describe it as something more than the familiar human habit of saving mental effort: the decision-maker "no longer constructs an answer, but adopts one generated by an external system."^[1] They propose that it is especially likely under time pressure, with complex tasks, when the person knows little about the subject, when trust in AI is high, or when the person wants to avoid effort. Users may even come to believe that the AI's reasoning is their own.^[1]

Distinction from related concepts

Shaw and Nave set cognitive surrender apart from two older ideas:

The authors write that "whereas cognitive offloading is a strategic delegation of deliberation, using a tool to aid one's own reasoning, cognitive surrender is an uncritical abdication of reasoning itself," and that where automation bias concerns specific errors, cognitive surrender "describes a broader disposition of epistemic dependence."^[1:2] The cognitive offloading framework they contrast with is the one set out by Evan Risko and Sam Gilbert in 2016.^[1][6]

Tri-System Theory

Dual-process theories of reasoning distinguish fast, intuitive, often emotional processing (System 1) from slow, deliberate, analytical reasoning (System 2).^[1] Shaw and Nave argue that these theories assume all cognition happens inside the brain, an assumption that no longer fits people who routinely ask algorithms and chatbots to interpret information and make judgments. Their Tri-System Theory adds a System 3, which they characterise as external, automated, data-driven, and dynamic, and which can either supplement or supplant a person's internal reasoning.^[1:3]

The theory describes several "routes" a judgment can take:^[1]

The authors also describe mixed routes, such as checking an AI answer before adopting it, overriding it, or rationalising an AI-supplied answer after the fact.^[1]

Experimental evidence

Design

Shaw and Nave tested the theory in three preregistered experiments with 1,372 participants and 9,593 individual trials, run with participants at a university behavioural laboratory and online via the Prolific platform.^[1:4][1:8] Participants solved seven items adapted from the Cognitive Reflection Test (CRT), a set of puzzles designed so that the first answer that comes to mind is wrong and the correct answer requires stopping to think.^[1] In the AI-assisted conditions, a chatbot built on OpenAI's GPT-4o model was embedded in the survey. Participants could consult it as often as they liked or not at all. Hidden instructions made it give either the correct answer or the tempting intuitive wrong answer to each item, at random, along with a short explanation.^[1:4] Participants were always free to ignore its advice.^[1]

Main findings

In the first study, 359 participants were randomly assigned to work either without AI or with access to the chatbot.^[1] Those with access consulted it on just over half of the items. When they consulted it, they followed its answer 92.7% of the time when it was correct and 79.8% of the time when it was wrong.^[1:5] As a result, their accuracy depended heavily on the AI's: participants without AI answered 45.8% of items correctly, compared with 71.0% on items where the AI was accurate and 31.5% on items where it was faulty.^[1:9] Access to the chatbot also raised participants' confidence in their answers by 11.7 percentage points, even though about half of its answers were wrong.^[1:6]

Pooling all three studies, the authors classified what happened on trials where participants consulted the AI and it was wrong. On 73.2% of these trials participants adopted the wrong answer (cognitive surrender); on 19.7% they rejected it and answered correctly, which the authors counted as cognitive offloading; and on 7.1% they rejected it but still answered incorrectly.^[1:10]

Situational factors

The second and third studies tested whether circumstances change the pattern. In the second study (485 participants), a 30-second limit per item reduced accuracy overall. When the AI was right, it offset some of that cost, but the gap between accurate-AI and faulty-AI trials remained.^[1:8] Under time pressure, correct overrides of faulty AI fell to 6.2% of such trials.^[1:10]

The third study (450 participants) paid a bonus of $0.20, plus a lottery entry, for each correct answer, and told participants immediately after each item whether they had been right.^[1] This combination more than doubled the rate at which participants rejected faulty AI advice, from 20.0% to 42.3%.^[1:11] In the pooled data, incentives and feedback raised correct overrides to 37.1% and lowered surrender to 57.9%.^[1:10] The authors conclude that surrender can be reduced but was not eliminated by any of the conditions they tested.^[1]

Individual differences

Susceptibility varied between people. Participants who reported more trust in AI were more likely to follow faulty advice. Those who scored higher on need for cognition, a measure of how much a person enjoys effortful thinking, and on a short test of fluid intelligence, were more likely to catch and correct the AI's errors.^[1:12][7]

Later research and use of the term

Within months, other researchers were using the term to interpret their own findings.

• Education. A May 2026 preprint by Sina Rismanchian and colleagues analysed 3.2 million interactions on the ALEKS online learning platform. After the release of ChatGPT, college students spent 26.9% less time on math word problems that AI can easily solve. The drop disappeared in proctored settings, and performance on proctored retention questions fell. The authors described the pattern as "the population-level indicator of cognitive surrender."^[3:1]
• Suspending judgment. A July 2026 preprint by Chiara Marcoccia, Walter Quattrociocchi, and Valerio Capraro called cognitive surrender the concept closest to their own concern, but noted that Shaw and Nave had measured only whether people ended up right or wrong, not whether they declined to answer.^[4:1] Across five experiments with 3,132 participants in which AI advice was deliberately wrong, access to AI almost eliminated participants' willingness to say "I don't know". They answered more questions but were correct about a third as often, while their confidence nearly doubled.^[4:2]
• Oversight of AI agents. In an August 2026 position paper on human oversight of autonomous AI agents, Margaret Mitchell, Avijit Ghosh, and Samir Passi listed cognitive surrender alongside "cognitive dependence" and "cognitive debt" as recent terms for the erosion of human skills through AI use. Among their proposed safeguards is rotating overseers between AI-assisted and unassisted versions of the same task to prevent "the cognitive surrender from prolonged exposure to agentic AI."^[5:1]

Reception and limitations

Writing for Gizmodo in April 2026, Mike Pearl called cognitive surrender "a new and useful term." He also advised readers to keep psychology's replication crisis in mind before treating any single experiment's results as settled, while adding that he had not noticed problems with the authors' work.^[2:1]

Shaw and Nave themselves note several limitations. The experiments took place in controlled settings that may not reflect everyday AI use. All three relied on a single type of task, the Cognitive Reflection Test. And each captured only one session, so they cannot show how trust and surrender change as people use AI repeatedly over time.^[1:13] By design, about half of the chatbot's answers were wrong, and each wrong answer was the item's tempting intuitive error.^[1:4][1:6] The authors call for field studies in areas such as financial apps and health platforms, tests with other kinds of reasoning, and research on further moderators such as personal accountability and technological skill.^[1:13]

The authors also caution that surrender is not inherently irrational: in some domains, deferring to a statistically superior system "may be adaptive or even optimal."^[1:7] In their experiments, participants who consulted an accurate AI did much better than those working alone.^[1:9] Their proposed remedies aim to preserve these gains while encouraging users to check answers: interfaces that signal uncertainty or confidence, incentives and feedback that reward checking, and education in when to trust AI output.^[1]

Analysis: cognitive surrender and human agency

Cognitive surrender bears directly on human agency, understood as people's durable capacity to understand their situation, form aims, and act on them. The experimental findings suggest that the harm lies less in using AI than in losing track of whose judgment is being acted on. A person who offloads a calculation still understands and owns the decision. A person who surrenders adopts a conclusion they did not evaluate, often with more confidence than if they had reasoned alone.^[1:6][4:2] That combination of higher confidence and lower independent checking weakens the ability to notice when something has gone wrong, which is the capacity that human oversight of automated systems depends on.^[5:1]

The evidence also suggests that surrender responds to the environment, not only to individual traits. Time pressure made it worse; immediate feedback and incentives to be right made it better.^[1:10] This places part of the responsibility on the design of AI products and of the institutions that deploy them. Systems that give fluent, confident answers with no cue about their reliability, in settings that reward speed over accuracy, can be expected to encourage surrender. Systems that surface uncertainty and make checking easy may help keep users' own reasoning engaged. Whether such interventions work outside the laboratory, and how surrender develops over months or years of everyday AI use, remain open empirical questions. Field studies and longitudinal studies of the kind the originators call for could answer them.^[1:13]