Cognitive Surrender: The Lineage of Handing Judgment to AI, What Generative AI Changed, and the Current Evidence on Countermeasures

When the AI is wrong, people do worse than without it

In January 2026, Shaw and Nave at Wharton gave participants access to an AI chat assistant while they solved an adapted Cognitive Reflection Test (CRT), a set of problems that measures whether people can suppress an intuitive but wrong answer and reconsider. Whether the AI’s answer was correct or wrong was randomized for each trial through a prompt that participants could not see. In the first experiment (N=359), participants consulted the AI on about half of the trials (54.4% of the trials with a correct AI and 52.8% of those with a wrong AI). On the trials in which they consulted it, they followed the AI 92.7% of the time when it was correct and 79.8% of the time when it was wrong. As a result, accuracy was 45.8% in the condition without AI, 71.0% on trials with a correct AI, and 31.5% on trials with a wrong AI. When the AI erred, participants did worse than they did without it. Moreover, participants with AI access were more confident in their answers than those without it (77.0% versus 65.3%).

Shaw and Nave called this pattern cognitive surrender. The same pattern has been studied for nearly 50 years under other names. What generative AI changed is less the phenomenon itself than how widely it spreads and what follows from it, namely the loss of learning and skill.

Scope and method

The collection was run in academic mode and divided into three clusters. The first covered the lineage before generative AI (human factors, medical informatics, cognitive psychology, philosophy, and judgment research), the second covered concepts and evidence from the generative AI period (November 2022 onward), and the third covered evidence on countermeasures. We screened 116 candidates for duplicates, retractions, corrections, and peer-review status (source/review/cognitive-surrender/papers.md), and this note cites 64 of them. All numbers and causal claims were checked against the full text, or, where the full text was unavailable, against an author version, an accepted manuscript, or the abstract. For sources whose abstract was the only text available, we report only what the abstract states and attribute it to the authors. Preprints are marked as not yet peer reviewed in the text. The paper by Shaw and Nave that introduced the term is itself a preprint (SSRN and PsyArXiv) as of November 2026.

What cognitive surrender refers to

Shaw and Nave added a third system, System 3, to dual-process theory, which distinguishes fast intuition (System 1) from slow deliberation (System 2). System 3 is automated, data-driven reasoning that originates in algorithmic systems outside the human mind, and people incorporate it into their reasoning as a partner they consult. They call this framework Tri-System Theory.

On this basis, they define cognitive surrender as “the behavioral and motivational tendency to defer judgment, effort, and responsibility to System 3’s output, particularly when that output is delivered fluently, confidently, or with minimal friction.” The key point of the definition is that what people hand over includes not only the answer but also the responsibility for the judgment and the effort of thinking.

The term is distinguished from two existing concepts. The first is cognitive offloading (Risko and Gilbert 2016), the strategic delegation of a specific task to a tool, such as a calculator or GPS. In offloading, one’s own reasoning keeps running, and the tool supports it. In surrender, the user gives up the reasoning itself and adopts the AI’s answer as their own. The second is automation bias, the omission and commission errors that arise from relying on automated cues. Shaw and Nave state that automation bias concerns specific errors, whereas cognitive surrender describes a broader disposition of epistemic dependence, and that a user in surrender “does not just follow System 3: they stop deliberative thinking altogether.”

In the experiments, following a wrong AI and answering incorrectly was counted as surrender, and overriding a wrong AI and answering correctly was counted as offloading. Across the three experiments, among trials in which participants consulted an AI that was wrong, 73.2% showed surrender, 19.7% showed offloading, and 7.1% were failed overrides, in which participants rejected the AI but still answered incorrectly. Susceptibility differed across individuals: participants with higher trust in AI were more likely to follow a wrong AI (odds ratio 3.47), whereas those higher in need for cognition (the tendency to enjoy effortful thinking) and fluid intelligence were less likely to do so.

Relying on AI can therefore mean two different things: delegating a task while keeping one’s own reasoning, or handing over the reasoning itself. The consequences differ. The former raises efficiency, whereas the latter makes people perform below their own unaided level when the AI is wrong.

The lineage before generative AI

Aviation and human factors: complacency and automation bias

The phenomenon was first named in aviation. From the 1970s, NASA’s Aviation Safety Reporting System (ASRS) included “complacency” as a coding item for incident reports and defined it as “self-satisfaction that may result in non-vigilance based on an unjustified assumption of satisfactory system state” (as described by Parasuraman and Manzey 2010, citing Billings et al. 1976). Parasuraman, Molloy, and Singh (1993) brought this concept into the laboratory. Forty participants performed manual tracking and fuel management in a flight simulation while monitoring an automated system for failures, and their detection of automation failures was substantially worse when automation reliability was constant than when it varied.

On the side of decision aids, Mosier, Skitka, and colleagues formulated automation bias from the mid-1990s. Mosier et al. (1998) defined automation bias as omission and commission errors that result from using automated cues as a heuristic replacement for vigilant information seeking and processing, and they examined it with glass-cockpit pilots. Parasuraman and Riley (1997) called overreliance on automation “misuse” and paired it with “disuse,” the neglect or underuse of automation. Lee and See (2004) framed the problem as people failing to rely on automation appropriately, argued that trust guides reliance when complexity and unanticipated situations make full understanding impractical, and set appropriate reliance as the design goal.

Parasuraman and Manzey (2010) treated complacency and automation bias as “different manifestations of overlapping automation-induced phenomena” and integrated them with attention at the center. This review concluded that automation bias occurs in both naive and expert participants and cannot be prevented by training or by explicit instructions to verify. This conclusion carries directly into the debate on countermeasures in the generative AI period.

Clinical decision support

In medicine, the same phenomenon was studied with clinical decision support systems (CDSS). Goddard et al. (2012) conducted a systematic review that selected 74 studies from 13,821 records and noted that the concept of automation bias was first discussed in aviation before spreading to healthcare. In a meta-analysis of four controlled studies, when the CDSS gave erroneous advice, the risk of an incorrect decision was 26% higher than in the control groups (risk ratio 1.26). The mitigators identified were training, emphasizing user accountability, the position of advice on the screen, updated confidence levels attached to the output, and providing information rather than recommendations. However, the same review also noted that training more clearly reduces complacency than automation bias and that the effect of accountability differed across studies. Lyell and Coiera (2017) found that automation bias was associated with the cognitive load of the task and suggested strategies that reduce cognitive load.

In diagnostic imaging, the label “from an AI” can matter more than the content of the advice. Gaube et al. (2021) had physicians read chest X-rays with advice that was all written by human experts, part of which was labeled as coming from an AI. Radiologists rated the AI-labeled advice as lower in quality, whereas physicians with less task expertise did not. When the advice was inaccurate, diagnostic accuracy was significantly worse regardless of the labeled source. In Dratsch et al. (2023), 27 radiologists read mammograms, and when the suggestion of a purported AI was wrong, the proportion of correct ratings fell from 79.7% to 19.8% among inexperienced readers, from 81.3% to 24.8% among moderately experienced readers, and from 82.3% to 45.5% even among very experienced readers. Experience weakened this effect but did not prevent it.

Cognitive psychology: knowledge held outside and the illusion of knowing

The cognitive psychology lineage has examined what happens when memory and processing are delegated to external tools. Sparrow et al. (2011) reported that when people expect to have future access to information, they remember where to find it rather than the information itself (the Google effect). However, one experiment in this paper was not reproduced in a replication study (Hesselmann 2020), and the robustness of the effect remains uncertain. Fisher et al. (2015) reported across nine experiments that people who searched the Internet for explanations mistook access to information for their own understanding and overestimated their knowledge. Dahmani and Bohbot (2020) reported that people who habitually used GPS had weaker spatial memory when navigating without it, and that among 13 participants retested three years later, greater GPS use during the interval was associated with a steeper decline (the longitudinal sample was small).

These studies indicate that processing delegated to a tool tends not to remain in the person, and that people rarely notice that it has not remained. In the experiment by Shaw and Nave, confidence did not fall after participants followed a wrong AI, which is consistent with this lineage.

On the side of dual-process theory, Frederick’s (2005) CRT became widely used as a measure of the ability or disposition to reflect on a question and resist reporting the first response that comes to mind. Shaw and Nave adapted the CRT and placed the AI’s answer in the position of the “first response,” so that their task tests whether people can override it.

Philosophy: the extended mind and epistemic vigilance

In philosophy, a position that does not treat dependence on tools as bad in itself came first. Clark and Chalmers (1998) argued that when an external resource functions as part of a cognitive process, the mind extends to that resource (the extended mind). Shaw and Nave also state that they built Tri-System Theory on this argument, on cognitive offloading, and on automation bias. On the other hand, Sperber et al. (2010) argued that, because humans depend massively on communication with others, they have a suite of cognitive mechanisms for evaluating sources and content, which they called epistemic vigilance. Cognitive surrender can be read as a state in which this vigilance stops working in front of an AI source. The line between dependence itself and dependence without scrutiny had already been drawn on the philosophical side.

Judgment research: avoiding and following algorithms

Judgment and decision-making research also knows the opposite phenomenon. Dietvorst et al. (2015) reported that after seeing an algorithm err, people avoid it even when it outperforms human forecasters (algorithm aversion). Both aversion and surrender are mismatches between an AI’s actual capability and the degree of reliance, and they correspond to the two directions in which people can depart from Lee and See’s appropriate reliance.

Skill degradation: the ironies of automation

Bainbridge (1983) argued that by mechanizing routine tasks and leaving exception handling to the human, automation deprives people of routine opportunities to practice their judgment (the ironies of automation). This argument is still cited, for example at the opening of the paper by Lee et al. (2025) on generative AI and critical thinking. Casner et al. (2014) had 16 airline pilots fly a simulator and found that their manual control and instrument scanning skills were mostly intact, whereas cognitive tasks such as tracking the aircraft’s position without a map display and deciding which navigational step comes next showed more frequent problems. In this study, the degradation was more visible in cognitive work than in manual skill, and cognitive work of this kind is what generative AI takes over.

What generative AI changed

Who delegates, what, and in what form

Compared with this lineage, generative AI changed the shape of the problem in three respects.

The first concerns who delegates. Research on automation bias studied trained professionals, such as pilots, physicians, and air traffic controllers, using specific tools. Generative AI is used by students, knowledge workers, and the general public for almost any question.

The second concerns what is delegated. Earlier automation returned narrow outputs such as alarms, diagnostic candidates, and suggested routes. Generative AI returns reasoning, explanations, text, and code, and enters the thinking process itself as what Shaw and Nave call System 3.

The third concerns how the output arrives. The output of generative AI is fluent, confident, and adapted to the user. This is why the definition by Shaw and Nave includes the condition that the output is “delivered fluently, confidently, or with minimal friction.” Zhou et al. (2024) reported that language models are reluctant to express uncertainty, that an average of 47% of the answers in which they expressed confidence were wrong, and that users relied heavily on generated answers whether or not they were marked by certainty. Steyvers et al. (2025) showed that people who read a model’s default explanations overestimated the accuracy of its answers, and that longer explanations increased confidence even when the extra length did not improve accuracy. Cheng et al. (2026) reported that 11 models affirmed users’ actions about 50% more often than humans did, and that in two preregistered experiments (N=1,604), sycophantic responses were rated as higher in quality, were trusted more, and made participants more willing to return to the model. The mechanisms that produce sycophancy are covered in Why Do LLMs Become Sycophantic? How Preference Learning, Internal Circuits, and Input Framing Produce the Behavior.

Following errors: evidence from the generative AI period

Evidence from the generative AI period reproduces the pattern of automation bias across a wider range of tasks.

Dell’Acqua et al. (2026) conducted a preregistered field experiment with 758 consultants at Boston Consulting Group. On 18 tasks within the frontier of AI capabilities, consultants using AI completed 12.2% more tasks, 25.1% faster, and with higher quality. On a task selected to lie outside that frontier, however, about 84.5% of the group without AI produced correct solutions, compared with 60% and 70.6% in the two AI conditions, an average decrease of 19 percentage points. The authors called the uneven boundary between what AI does well and badly, which does not match apparent task difficulty, the “jagged technological frontier.”

Spatharioti et al. (2025) showed that in product comparison tasks using LLM-based search, users relied on incorrect information when the LLM erred. Perry et al. (2023) reported that in security programming tasks with 47 participants, those with access to an AI assistant wrote less secure code and were more likely to believe that their code was secure. Marcoccia et al. (2026, not yet peer reviewed) conducted five experiments (N=3,132). They reported that merely having access to AI advice nearly eliminated participants’ willingness to suspend judgment by saying “I don’t know.” On questions engineered so that the advice was usually wrong, participants with AI access were correct about a third as often as without AI, while their confidence nearly doubled. The rate of suspending judgment thus offers a second indicator of surrender that the binary of following or overriding the AI cannot capture.

Generative AI also strengthens illusions on the side of metacognition, the ability to estimate one’s own performance. Fernandes et al. (2026) reported that 246 participants who solved 20 logical reasoning problems from the LSAT with ChatGPT scored about three points higher than a norm population but overestimated their own score by about four points. The Dunning–Kruger effect usually observed in this task, in which low performers overestimate themselves the most, disappeared with AI use, and participants with more technical knowledge of AI judged their own performance less accurately (replicated in a second study with N=452). Lee et al. (2025) analyzed 936 examples from 319 knowledge workers and found that higher confidence in generative AI was associated with less critical thinking, whereas higher self-confidence was associated with more. They also found that the content of critical thinking shifted from information gathering to verification and from task execution to response integration and task stewardship (the study is a self-report survey). Gerlich (2025) states in the abstract that, in surveys and interviews with 666 participants, a negative correlation between frequent AI tool use and critical thinking was mediated by increased cognitive offloading (the study is correlational, and a correction of an error in a table was later published).

A meta-analysis answers whether combining humans and AI improves performance. Vaccaro et al. (2024) synthesized 370 effect sizes from 106 experiments published between January 2020 and June 2023 and found that, on average, human–AI combinations performed worse than the better of humans or AI alone (Hedges’ g = −0.23), with losses in decision tasks and larger gains in content creation tasks. Passi and Vorvoreanu (2022) noted that calls for human oversight make users “the last line of defense against AI failures,” and the results of Vaccaro et al. indicate that, on average, this line of defense does not hold.

Beyond following errors: fewer opportunities to learn, less skill

What came to the foreground in the generative AI period is less the error in the immediate judgment than the loss of learning and skill through repeated delegation.

In an educational field experiment, Bastani et al. (2025) had nearly 1,000 high school students in Turkey practice mathematics. Students with access to a standard GPT-4 interface improved their practice grades by 48%, but on an exam taken without AI they scored 17% lower than students who never had access. In the GPT Tutor arm, designed to give hints rather than answers, practice grades rose by 127%, and the drop on the exam largely disappeared (although the arm did not outperform the control). Fan et al. (2025) showed that students using ChatGPT outperformed the other groups in essay score improvement but did not differ in knowledge gain or transfer, and they named the state of handing over the self-regulation of learning to AI metacognitive laziness.

Studies of skill formation and large-scale behavioral logs point in the same direction. In a randomized experiment in which developers learned a new programming library (n=52), Shen and Tamkin (2026, not yet peer reviewed; the authors are affiliated with Anthropic) found that the group using AI scored 17% lower (two grade points) on a comprehension quiz, without a statistically significant reduction in completion time. Participants who fully delegated code generation to AI learned less, whereas three interaction patterns that involved cognitive engagement, such as asking for explanations or asking only conceptual questions, preserved learning. Rismanchian et al. (2026, not yet peer reviewed) analyzed 3.2 million learning interactions on the ALEKS learning system, and the authors include employees of McGraw Hill, which provides ALEKS. They found that after the release of ChatGPT, time spent on problems that AI can solve declined, a decline absent under proctoring, and that the odds of a correct response on proctored retention items fell by 25%. These effects on learning are covered in more detail in the literature map on generative AI and learning, Does Early Use of Generative AI Inhibit the Formation of Thought? A Literature Map of Cognitive Offloading and Learning.

In neurophysiological measurement, Kosmyna et al. (2025, not yet peer reviewed) recorded EEG while 54 participants wrote essays and reported that the LLM group showed the weakest neural coupling and low ownership of their essays, which they called cognitive debt. However, Stankovic et al. (2025) published a comment that raised concerns about the small sample, the reproducibility of the analyses, the EEG methods, and inconsistencies in the reported results, so this study is best treated as supporting evidence only.

Observational evidence has also appeared for professional skill. Budzyń et al. (2025) compared the periods before and after AI support was introduced for colonoscopy at four centers in Poland and reported that the adenoma detection rate of standard colonoscopy without AI fell from 28.4% (226/795) to 22.4% (145/648) (values from the accepted manuscript; as an observational study, it does not establish causation). This is a present-day instance of Bainbridge’s irony: once experts become used to AI, their performance drops when the AI is absent.

At the collective level, Doshi and Hauser (2024) reported that stories written with access to generative AI ideas were individually evaluated as more creative, but became more similar to each other. In this story-writing task, when individuals relied on AI ideas, the diversity of the whole decreased.

How the relevant academic fields have shifted

Rearranging this lineage by field shows that the center of research on the same phenomenon has moved over time.

The shift has three strands. First, the object of concern moved from the safety of professionals to the learning and thinking of the general public. Human factors asked whether accidents could be prevented, whereas the learning sciences and cognitive psychology in the generative AI period ask whether the capacity to think and the opportunities to learn remain. Second, the theoretical framing moved from models of attention and trust to extensions of dual-process theory. Chiriatti et al. (2024) proposed System 0, which places interaction with AI before thinking, and Shaw and Nave proposed System 3 as a third system of reasoning; both began to treat AI as a participant in reasoning. Third, what is measured changed. Earlier studies measured error rates while AI was present, whereas studies in the generative AI period measure what remains after AI is removed (exams, retention tests, and examination performance without AI). This change overlaps with the trend summarized in Research Trends in HCI since Generative AI Became Usable (2023–2026): Many Systems Built, and the Checking Work Left to People, in which HCI after generative AI began to address the shift from making to checking.

However, the connections between fields remain weak. Shaw and Nave cite automation bias and Mosier and Skitka, but many HCI papers from the generative AI period do not examine, as a premise for their own countermeasures, the three decades of findings from human factors (that training does not prevent automation bias and that the effect of accountability is inconsistent). Xu et al. (2026, not yet peer reviewed) automatically classified 1,223 AI–HCI papers from 2023 to early 2026 and reported that research assuming a frictionless design paradigm still accounted for 67.3% in 2026, while research defending human epistemic sovereignty rose to 19.1% in 2025 and then fell to 13.1% in early 2026 (the classification was performed automatically by a language model).

Countermeasures: what works and what does not

Evidence by countermeasure

The intervention studies differ in task, participants, and AI accuracy, so their numbers cannot be compared directly. The following table lists only the direction of effects and the conditions.

Five principles that follow from the table

The countermeasures that work share several properties.

First, interventions that make AI output easier to believe tend to strengthen surrender. Explanations, longer explanations, a confident tone, and sycophancy all make output easier to accept. Romeo and Conti (2025) systematically reviewed 35 peer-reviewed studies from January 2015 to April 2025 and state in the abstract that explainable AI and transparency mechanisms may increase perceived acceptability but are often insufficient to improve decision accuracy or mitigate automation bias. Explanations help only when they serve as a means of verification; when they serve as a means of persuasion, they backfire.

Second, effective interventions reactivate the user’s deliberation (System 2) locally, where it is needed. Item-level feedback, cues where errors are likely, and showing sources and inconsistencies all give users a signal that a particular point is worth checking. Shaw and Nave explain the effect of incentives and feedback not as a blunt suppression of AI help but as a targeted reactivation of System 2, in which users pause, check, and depart from the AI when it conflicts with available cues. General warnings may fail because the signal is not local and does not tell users when to check.

Third, effective interventions have costs. In Buçinca et al. (2021), the designs that reduced overreliance the most were liked the least. In Fogliato et al. (2022), the procedure of recording one’s own judgment first reduced agreement with correct AI along with agreement with wrong AI, and it lowered the perceived usefulness of the AI. Raees et al. (2026) state in the abstract of their review that user engagement benefits appropriate reliance but comes at a cost. Whether disliked designs are chosen and kept in real products has hardly been studied.

Fourth, the effects are concentrated among people with high motivation to think and more experience. Cognitive forcing and partial explanations benefited people high in need for cognition more (Buçinca et al. 2021; de Jong et al. 2025), and in the experiments by Shaw and Nave, people higher in need for cognition and fluid intelligence were less likely to surrender. Countermeasures therefore tend to protect those who are already less likely to surrender while reaching those who are more likely to surrender less often. As Fernandes et al. (2026) showed, technical knowledge of AI does not help people estimate their own performance correctly either.

Fifth, protecting learning and skill requires measuring performance after AI is removed rather than performance with AI. Bastani et al., Fan et al., Shen and Tamkin, Rismanchian et al., and Budzyń et al. all detected harm through performance without AI. Assessment that looks only at performance with AI mistakes surrender for success. Corbin et al. (2025) distinguished discursive changes, which only communicate rules that students remain free to ignore, from structural changes to the mechanics of assessment, and argued that the former cannot protect validity.

The current state of regulation and policy

Regulation and standards now name automation bias. Article 14(4)(b) of the EU AI Act (Regulation 2024/1689) requires that people assigned to oversee high-risk AI systems be enabled “to remain aware of the possible tendency of automatically relying or over-relying on the output produced by a high-risk AI system (automation bias).” The provision applies in particular to systems that provide information or recommendations for decisions taken by natural persons. The Generative AI Profile of the U.S. National Institute of Standards and Technology (NIST AI 600-1) describes automation bias as an example of “excessive deference to automated systems” and states that it can exacerbate other risks such as confabulation, bias, and homogenization.

However, the premise that human oversight solves the problem faces a strong objection. Green (2022) surveyed 41 policies that require human oversight of government algorithms and found two flaws: evidence suggests that people are unable to perform the desired oversight functions, and as a result, oversight policies legitimize the use of faulty algorithms. He proposed a shift from individual human oversight to institutional oversight. Taken together with the conclusion of Parasuraman and Manzey (2010) that training and instructions do not prevent automation bias, it is difficult to expect that a legal requirement to “remain aware” will by itself prevent surrender.

In education, Suazo Galdames et al. (2026) systematically reviewed 26 empirical studies in higher education and concluded that uncritical reliance on AI output was frequently associated with cognitive offloading, automation bias, superficial information processing, and weakened evaluative judgment, while the consequences depended on pedagogical design, AI literacy, and assessment design.

Synthesis: what is known and what is not

Cognitive surrender is not a new phenomenon. A phenomenon studied for more than 30 years in aviation and medicine as automation complacency and automation bias has, with generative AI, spread from professionals to the general public, from narrow alarms to reasoning itself, and from errors in the immediate decision to losses of learning and skill. The contribution of Shaw and Nave is to recast this spread in the language of dual-process theory and to make delegation (offloading) and abdication (surrender) measurable as distinct outcomes.

The evidence on countermeasures supports the following. Making AI output easier to believe (explanations, fluency, general warnings) does not reduce surrender. The interventions with evidence of effect are local cues where errors are likely, showing sources and inconsistencies, first-person expressions of uncertainty, item-level feedback with incentives, and assessment that measures performance without AI. Procedures in which people record their own judgment first reduce reliance on correct AI as well, and designs in which the AI argues back still have little evidence. Every countermeasure carries costs, namely lower user preference, higher load, and uneven benefits across people, and none of them overturns the human factors conclusion that training and regulation alone do not prevent automation bias.

Unresolved disagreements

Gaps in the evidence

• The paper that introduced the term cognitive surrender is itself not yet peer reviewed, and there is no independent replication yet.
• Studies of the long-term effects of surrender (skill and judgment after months or years of delegation) are limited to an observational study (Budzyń et al. 2025) and behavioral logs (Rismanchian et al. 2026); we found no randomized longitudinal study.
• Most intervention studies use short online experiments and simple tasks (mazes, nutrition, fractions, multiple-choice questions), and Raees et al. (2026) also note that realistic decision scenarios are rarely replicated.
• Whether disliked but effective designs remain in use in products where users can choose has not been measured.
• We found no evidence on surrender in domains such as design, where there is no single correct answer and it is hard to see whose judgment was adopted.
• We found no evidence from Japanese-speaking settings in this search.

Unverified items

• A correction notice was published for Budzyń et al. (2025) (Lancet Gastroenterology & Hepatology, 10.1016/S2468-1253(25)00294-8), but its content could not be retrieved. The numbers in this note are those of the authors’ accepted manuscript.
• A correction notice for an error in a table was also published for Gerlich (2025) (Societies 15(9): 252), but which values changed could not be retrieved. This note uses only the abstract.
• The numbers from Fernandes et al. (2026) were checked against the author version (arXiv). Differences from the published version were not checked.
• The full texts of Chiang et al. (2024), de Jong et al. (2025), Romeo and Conti (2025), Raees et al. (2026), Corbin et al. (2025), and Gerlich (2025) could not be retrieved, and these are described within the scope of their abstracts.
• The same applies, in the lineage section, to Parasuraman et al. (1993), Parasuraman and Riley (1997), Mosier et al. (1998), Lee and See (2004), Casner et al. (2014), Frederick (2005), Sparrow et al. (2011), and Fisher et al. (2015).
• Neither the full text nor the abstract of Bainbridge (1983), Risko and Gilbert (2016), or Chiriatti et al. (2024) could be retrieved; they are described through the accounts in the full texts of Lee et al. (2025) and Shaw and Nave (2026).

References

All accessed on 2026-11-06. See source/review/cognitive-surrender/papers.md for the ledger and verification records.

Origin of the term and concepts from the generative AI period

• Shaw, S. D., & Nave, G. (2026). Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender. SSRN Working Paper 6097646 / PsyArXiv. https://doi.org/10.2139/ssrn.6097646 ／ https://doi.org/10.31234/osf.io/yk25n_v1 (preprint, not yet peer reviewed)
• Chiriatti, M., Ganapini, M., Panai, E., Ubiali, M., & Riva, G. (2024). The case for human–AI interaction as system 0 thinking. Nature Human Behaviour, 8(10), 1829–1830. https://doi.org/10.1038/s41562-024-01995-5
• Xu, K., Shen, Y., Yan, L., & Ren, Y. (2026). Cognitive Agency Surrender: Defending Epistemic Sovereignty via Scaffolded AI Friction. arXiv:2603.21735. https://arxiv.org/abs/2603.21735 (preprint)

The lineage before generative AI

• Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098%2883%2990046-8
• Parasuraman, R., Molloy, R., & Singh, I. L. (1993). Performance consequences of automation-induced “complacency”. International Journal of Aviation Psychology, 3(1), 1–23. https://doi.org/10.1207/s15327108ijap0301_1
• Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
• Mosier, K. L., Skitka, L. J., Heers, S., & Burdick, M. (1998). Automation bias: Decision making and performance in high-tech cockpits. International Journal of Aviation Psychology, 8(1), 47–63. https://doi.org/10.1207/s15327108ijap0801_3
• Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50.30392
• Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055
• Goddard, K., Roudsari, A., & Wyatt, J. C. (2012). Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association, 19(1), 121–127. https://doi.org/10.1136/amiajnl-2011-000089
• Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W. (2014). The retention of manual flying skills in the automated cockpit. Human Factors, 56(8), 1506–1516. https://doi.org/10.1177/0018720814535628
• Lyell, D., & Coiera, E. (2017). Automation bias and verification complexity: a systematic review. Journal of the American Medical Informatics Association, 24(2), 423–431. https://doi.org/10.1093/jamia/ocw105
• Gaube, S., Suresh, H., Raue, M., et al. (2021). Do as AI say: susceptibility in deployment of clinical decision-aids. npj Digital Medicine, 4, 31. https://doi.org/10.1038/s41746-021-00385-9
• Frederick, S. (2005). Cognitive reflection and decision making. Journal of Economic Perspectives, 19(4), 25–42. https://doi.org/10.1257/089533005775196732
• Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745
• Hesselmann, G. (2020). No conclusive evidence that difficult general knowledge questions cause a “Google Stroop effect”. A replication study. PeerJ, 8, e10325. https://doi.org/10.7717/peerj.10325
• Fisher, M., Goddu, M. K., & Keil, F. C. (2015). Searching for explanations: How the Internet inflates estimates of internal knowledge. Journal of Experimental Psychology: General, 144(3), 674–687. https://doi.org/10.1037/xge0000070
• Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
• Dahmani, L., & Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports, 10, 6310. https://doi.org/10.1038/s41598-020-62877-0
• Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7
• Sperber, D., Clément, F., Heintz, C., Mascaro, O., Mercier, H., Origgi, G., & Wilson, D. (2010). Epistemic vigilance. Mind & Language, 25(4), 359–393. https://doi.org/10.1111/j.1468-0017.2010.01394.x
• Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033

Evidence from the generative AI period

• Dell’Acqua, F., McFowland III, E., Mollick, E., et al. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403–423. https://doi.org/10.1287/orsc.2025.21838
• Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI ‘25. https://doi.org/10.1145/3706598.3713778
• Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006
• Steyvers, M., Tejeda, H., Kumar, A., et al. (2025). What large language models know and what people think they know. Nature Machine Intelligence, 7, 221–231. https://doi.org/10.1038/s42256-024-00976-7
• Fernandes, D., Villa, S., Nicholls, S., et al. (2026). AI makes you smarter but none the wiser: The disconnect between performance and metacognition. Computers in Human Behavior, 175, 108779. https://doi.org/10.1016/j.chb.2025.108779
• Marcoccia, C., Quattrociocchi, W., & Capraro, V. (2026). AI advice suppresses people’s willingness to say “I don’t know”, even when the advice is wrong and accuracy is incentivized. arXiv:2607.13562. https://arxiv.org/abs/2607.13562 (preprint)
• Spatharioti, S. E., Rothschild, D. M., Goldstein, D. G., & Hofman, J. M. (2025). Effects of LLM-based Search on Decision Making: Speed, Accuracy, and Overreliance. CHI ‘25. https://doi.org/10.1145/3706598.3714082
• Perry, N., Srivastava, M., Kumar, D., & Boneh, D. (2023). Do Users Write More Insecure Code with AI Assistants? CCS ‘23, 2785–2799. https://doi.org/10.1145/3576915.3623157
• Zhou, K., Hwang, J. D., Ren, X., & Sap, M. (2024). Relying on the Unreliable: The Impact of Language Models’ Reluctance to Express Uncertainty. ACL 2024, 3623–3643. https://doi.org/10.18653/v1/2024.acl-long.198
• Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352. https://doi.org/10.1126/science.aec8352
• Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290
• Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293–2303. https://doi.org/10.1038/s41562-024-02024-1
• Budzyń, K., Romańczyk, M., Kitala, D., et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology, 10(10), 896–903. https://doi.org/10.1016/S2468-1253%2825%2900133-5
• Dratsch, T., Chen, X., Rezazade Mehrizi, M., et al. (2023). Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance. Radiology, 307(4), e222176. https://doi.org/10.1148/radiol.222176
• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
• Fan, Y., Tang, L., Le, H., et al. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
• Kosmyna, N., Hauptmann, E., Yuan, Y. T., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872. https://arxiv.org/abs/2506.08872 (preprint)
• Stankovic, M., Hirche, E., Kollatzsch, S., & Doetsch, J. N. (2025). Comment on: Your Brain on ChatGPT. arXiv:2601.00856. https://arxiv.org/abs/2601.00856 (preprint)
• Shen, J. H., & Tamkin, A. (2026). How AI Impacts Skill Formation. arXiv:2601.20245. https://arxiv.org/abs/2601.20245 (preprint)
• Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E. (2026). Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build. arXiv:2605.21629. https://arxiv.org/abs/2605.21629 (preprint)

Evidence on countermeasures and regulation

• Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), 188. https://doi.org/10.1145/3449287
• Bansal, G., Wu, T., Zhou, J., et al. (2021). Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. CHI ‘21. https://doi.org/10.1145/3411764.3445717
• Vasconcelos, H., Jörke, M., Grunde-McLaughlin, M., Gerstenberg, T., Bernstein, M. S., & Krishna, R. (2023). Explanations Can Reduce Overreliance on AI Systems During Decision-Making. Proceedings of the ACM on Human-Computer Interaction, 7(CSCW1), 129. https://doi.org/10.1145/3579605
• Kim, S. S. Y., Liao, Q. V., Vorvoreanu, M., Ballard, S., & Vaughan, J. W. (2024). “I’m Not Sure, But…”: Examining the Impact of Large Language Models’ Uncertainty Expression on User Reliance and Trust. FAccT ‘24. https://doi.org/10.1145/3630106.3658941
• Kim, S. S. Y., Vaughan, J. W., Liao, Q. V., Lombrozo, T., & Russakovsky, O. (2025). Fostering Appropriate Reliance on Large Language Models: The Role of Explanations, Sources, and Inconsistencies. CHI ‘25. https://doi.org/10.1145/3706598.3714020
• Fogliato, R., Chappidi, S., Lungren, M., et al. (2022). Who Goes First? Influences of Human-AI Workflow on Decision Making in Clinical Imaging. FAccT ‘22. https://doi.org/10.1145/3531146.3533193
• Chiang, C.-W., Lu, Z., Li, Z., & Yin, M. (2024). Enhancing AI-Assisted Group Decision Making through LLM-Powered Devil’s Advocate. IUI ‘24. https://doi.org/10.1145/3640543.3645199
• de Jong, S., Paananen, V., Tag, B., & van Berkel, N. (2025). Cognitive Forcing for Better Decision-Making: Reducing Overreliance on AI Systems Through Partial Explanations. Proceedings of the ACM on Human-Computer Interaction, 9(2). https://doi.org/10.1145/3710946
• Vejandla, S., Ray, A., Sikstrom, L., Ratto, M., Zaheer, J., & Maslej, M. M. (2025). Impacts of cognitive forcing and need for cognition on biased AI-assisted decision making about mental health emergencies. Scientific Reports, 16, 818. https://doi.org/10.1038/s41598-025-30506-3
• Sarkar, A. (2024). AI Should Challenge, Not Obey. Communications of the ACM, 67(10), 18–21. https://doi.org/10.1145/3649404
• Tankelevitch, L., Kewenig, V., Simkute, A., et al. (2024). The Metacognitive Demands and Opportunities of Generative AI. CHI ‘24. https://doi.org/10.1145/3613904.3642902
• Maier, S., Schwabe, K., Schneider, M., & Feuerriegel, S. (2026). Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants. arXiv:2609.20143. https://arxiv.org/abs/2609.20143 (preprint)
• Schemmer, M., Kühl, N., Benz, C., Bartos, A., & Satzger, G. (2023). Appropriate Reliance on AI Advice: Conceptualization and the Effect of Explanations. IUI ‘23. https://doi.org/10.1145/3581641.3584066
• Passi, S., & Vorvoreanu, M. (2022). Overreliance on AI: Literature Review. Microsoft Research. https://www.microsoft.com/en-us/research/wp-content/uploads/2022/06/Aether-Overreliance-on-AI-Review-Final-6.21.22.pdf
• Romeo, G., & Conti, D. (2025). Exploring automation bias in human–AI collaboration: a review and implications for explainable AI. AI & Society, 41(1), 259–278. https://doi.org/10.1007/s00146-025-02422-7
• Raees, M., Khan, V.-J., Lykourentzou, I., & Papangelis, K. (2026). Do People Appropriately Rely on AI-Advice? An Analytical Review of HCI Research on Human-AI Decision-Making. CHI ‘26. https://doi.org/10.1145/3772318.3791467
• Blanchard, S. J., Garvey, A. M., & O’Laughlin, L. (2026). Hallucinations in organization-backed AI advisors: Evidence about skepticism, verification, and reliance in goal-directed use. Current Opinion in Psychology, 73, 102427. https://doi.org/10.1016/j.copsyc.2026.102427 (author version: https://arxiv.org/abs/2606.23491 )
• Suazo Galdames, I. C., Santiesteban Velázquez, M., Saracostti, M., & Chaple Gil, A. M. (2026). Between cognitive offloading and critical autonomy: a systematic review of the epistemic implications of generative AI in higher education. Frontiers in Education, 11. https://doi.org/10.3389/feduc.2026.1878665
• Corbin, T., Dawson, P., & Liu, D. (2025). Talk is cheap: why structural assessment changes are needed for a time of GenAI. Assessment & Evaluation in Higher Education, 50(7), 1087–1097. https://doi.org/10.1080/02602938.2025.2503964
• Green, B. (2022). The flaws of policies requiring human oversight of government algorithms. Computer Law & Security Review, 45, 105681. https://doi.org/10.1016/j.clsr.2022.105681
• Regulation (EU) 2024/1689 of the European Parliament and of the Council (Artificial Intelligence Act), Article 14. Official Journal of the European Union, 2024-07-12. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
• NIST (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1). https://doi.org/10.6028/NIST.AI.600-1