CoChat
Sign Up
Cited by 28Artificial Intelligence in Healthcare and EducationColorectal Cancer Screening and DetectionClinical Reasoning and Diagnostic Skills

Enhancing human-AI collaboration: The case of colonoscopy

Open access
Introzzi, Luca, Zonca, Joshua, Cabitza, Federico, Cherubini, Paolo, Reverberi, Carlo·2024·Digestive and Liver Disease
DOI 10.1016/j.dld.2023.10.018
Open this paper in CoChat — free accountOpen PDF

Abstract

Diagnostic errors impact patient health and healthcare costs. Artificial Intelligence (AI) shows promise in mitigating this burden by supporting Medical Doctors in decision-making. However, the mere display of excellent or even superhuman performance by AI in specific tasks does not guarantee a positive impact on medical practice. Effective AI assistance should target the primary causes of human errors and foster effective collaborative decision-making with human experts who remain the ultimate decision-makers. In this narrative review, we apply these principles to the specific scenario of AI assistance during colonoscopy. By unraveling the neurocognitive foundations of the colonoscopy procedure, we identify multiple bottlenecks in perception, attention, and decision-making that contribute to diagnostic errors, shedding light on potential interventions to mitigate them. Furthermore, we explored how existing AI devices fare in clinical practice and whether they achieved an optimal integration with the human decision-maker. We argue that to foster optimal Human-AI collaboration, future research should expand our knowledge of factors influencing AI's impact, establish evidence-based cognitive models, and develop training programs based on them. These efforts will enhance human-AI collaboration, ultimately improving diagnostic accuracy and patient outcomes. The principles illuminated in this review hold more general value, extending their relevance to a wide array of medical procedures and beyond.

AI snapshot

Written from the paper’s full text.

TL;DR

This narrative review applies cognitive science to colonoscopy to explain perceptual, attentional, and decision-making bottlenecks that cause missed lesions, and argues that AI can help only if human-AI interaction is designed and trained around those bottlenecks. It reviews CADe/CADx evidence, finding strong support for CADe while warning that real-world negative results may reflect overreliance, underreliance, and poor transparency.

Key findings

  • Diagnostic errors are estimated as high as 10–15%; tandem colonoscopy studies show about 25% of adenomas missed and 12% of index colonoscopies false negative.
  • Cognitive bottlenecks include guided search, a useful field of view of roughly 2.5° centered on the fovea (equivalent to 4.4 cm width at 1 m), poor memory for sampled locations, inhibition of return (0.3 s is enough to tag a location), vigilance decrement, dual-task interference, miscalibration, and overconfidence.
  • CADe evidence: a recent review found CADe systems reduced adenoma miss rate by 65%; false positives are not rare (e.g., ∼27 per colonoscopy), but experts often recognize them without delaying the procedure; longer AI signal persistence (e.g., 200 ms) is associated with a higher likelihood of a true lesion.
  • CADx evidence: AI output substantially influenced both expert and non-expert endoscopists; all were more likely to accept AI advice when correct than when wrong, improving integrated human-AI decision accuracy especially among non-experts, with a smaller but significant effect for experts. Some studies found no significant CADx benefit.

Methods

  • Narrative review, not a systematic review or primary empirical study; it synthesizes cognitive science, human-AI interaction principles, and colonoscopy AI literature.
  • Data sources are cited randomized controlled trials, meta-analyses, tandem colonoscopy studies, and real-world implementation studies of CADe/CADx, with Tables 1–4 summarizing cognitive bottlenecks, trust factors, and study lists.
  • Main analysis is conceptual/theoretical integration: mapping cognitive bottlenecks to potential AI supports and reviewing empirical evidence for CADe and CADx.

Limitations

  • The paper states that existing insights remain preliminary and a comprehensive understanding of the cognitive bottlenecks leading to performance disparities is lacking; hypotheses from negative real-world studies must be formally tested.
  • The paper notes that real-world negative studies were retrospective and lack RCT rigor, and that factors such as baseline performance, attitudes toward technology, trust, understanding of AI operations, interaction protocols, and tool features remain to be explored.
  • Evident from the text: as a narrative review, it does not provide a systematic search strategy, formal risk-of-bias assessment, or new empirical data; conclusions rely on selected cited literature.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.dld.2023.10.018

Cites

128 of 128 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

58 of 127 cited sources checked · 13 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    Even in ideal conditions where all relevant information is available, humans deviate from optimal integration of opinions.

    “Even in these ideal conditions, where all relevant information is available, humans deviate from an optimal integration of opinions [50 , 51] .”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [50] Bahrami, 2010, Optimally interacting minds, Science source

      The Main Text says the dyads made Bayes-optimal use of the confidence information they communicated. It also says the model in which participants shared both the signal and its reliability outperformed the dyads, but the authors considered it unlikely that participants shared both and then used them suboptimally. The paper therefore does not establish the claimed deviation despite all relevant information being available. (LLM, conf=0.86)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [51] Bahrami, 2012, What failure in collective decision-making tells us about metacognition, Philos Trans R Soc B source

      The Discussion reports that verbally communicating pairs made counterproductive collective decisions when their perceptual sensitivities differed, even though the competence gap was conspicuously obvious. This supports deviation from optimal opinion integration, but it does not establish the claim's ‘ideal conditions’ qualifier that all relevant information was available; the paper instead examines how differences in evidence quality and modes of communication affect decisions. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • Premisecitation

    When target and background differ over multiple perceptual dimensions, search becomes intentional and guided by a target template.

    “Our search becomes guided by a target template: Targets are defined by a multidimensional combination of features that segregate them from irrelevant stimuli or the background [10] .”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [10] Wolfe, 2021, Guided Search 6.0: an updated model of visual search, Psychon Bull Rev source

      In “Search Templates and Hybrid Search” (page unavailable), Wolfe says target features can guide attention away from irrelevant items. But the paper calls the representation that guides search a “guiding template” and distinguishes it from the “target template,” which is used to identify a selected item. It also does not require differences across multiple dimensions for search to be guided. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
  • Premisecitation

    Guided search can overlook relevant but untargeted objects, producing inattentional blindness.

    “The guided search can even overlook relevant but untargeted objects, producing astonishing phenomena like inattentional blindness [11 , 12] .”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [11] Williams, 2021, The invisible breast cancer: experience does not protect against inattentional blindness to clinically relevant findings in radiology, Psychon Bull Rev source
    • cites [12] Simons, 1999, Gorillas in our midst: sustained inattentional blindness for dynamic events, Perception source
  • Interpretationcitation

    Over-alternation bias may contribute to observed increase in misses following detecting the first polyp during a colonoscopy.

    “This tendency may contribute to the observed increase in misses following detecting the first polyp during a colonoscopy [7 , 27-29] .”
    Quote in the paperNot a substantial claim1 of 4 sources checked
    • cites [7] Kaminski, 2020, Optimizing the quality of colorectal cancer screening worldwide, Gastroenterology source

      The ‘How to optimize lesion detection?’ section discusses adenoma detection rates, and the Journal Pre-proof passages discuss factors associated with missed adenomas. Searches for over-alternation bias, the first polyp, and misses after an initial detection found no passage addressing that proposed connection. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
    • cites [27] Leufkens, 2012, Factors influencing the miss rate of polyps in a back-to-back colonoscopy study, Endoscopy source
    • cites [28] Wang, 2013, Adenoma detection rate is necessary but insufficient for distinguishing high versus low endoscopist performance, Gastrointest Endosc source
    • cites [29] Aniwan, 2016, The secondary quality indicator to improve prediction of adenoma miss rate apart from adenoma detection rate, Off J Am Coll Gastroenterol ACG source
  • Premisecitation

    The human brain's processing of sensory information and decision-making did not evolve to detect lesions flawlessly during standard endoscopy conditions.

    “The human brain's processing of sensory information and human decision-making did not evolve to detect lesions flawlessly during standard endoscopy conditions [9] .”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] Wolfe, 2022, Normal blindness: when we look but fail to see, Trends Cogn Sci source

      The Abstract says that evolved visual processes are virtually guaranteed to miss some significant stimuli, including in medical image perception. The section “Missing the designated target” also describes decision processes that can lead to missed findings. The paper does not address lesion detection during standard endoscopy, so that setting-specific part of the claim is unsupported. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • Classified as probably not a substantial claim. 19
  • May combine two claims. 23

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Diagnostic errors are as high as 10-15%, impacting patient health and system costs.

    “Estimates evaluate diagnostic errors as high as 10-15%, impacting patient health and system costs [1][2][3] .”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites [1] Graber, 2013, The incidence of diagnostic error in medicine, BMJ Qual Saf source
    • cites [2] Gunderson, 2020, Prevalence of harmful diagnostic errors in hospitalised adults: a systematic review and meta-analysis, BMJ Qual Saf source
    • cites [3] Makary, 2016, Medical error—the third leading cause of death in the US, BMJ source

  • May combine two claims.

    Premisecitation

    AI exhibits strong pattern recognition capabilities, relying on different features and mechanisms than humans, thereby likely failing for reasons distinct from those affecting human performance.

    “Moreover, AI exhibits strong pattern recognition capabilities, relying on different features and mechanisms than humans, thereby likely failing for reasons distinct from those affecting human performance [40] .”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [40] Szegedy, 2014, Intriguing, properties of neural networks source

      The Abstract (p. 1) reports strong recognition performance and shows that hardly perceptible changes can make a neural network misclassify an image. The Introduction (p. 2) describes this as an intrinsic blind spot. These passages support the suggestion that network failures can differ from human perception, but the paper does not establish that AI generally relies on different features and mechanisms than humans. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Aggregating informed opinions improves effectiveness, promoting the wisdom of crowds.

    “extensive evidence has shown that aggregating informed opinions improves effectiveness [50 , 58] , promoting the wisdom of crowds [59 , 60] .”
    Quote in the paperNot a substantial claim3 of 4 sources checked
    • cites [50] Bahrami, 2010, Optimally interacting minds, Science source

      The Main Text reports that combining two observers’ judgments can improve perceptual sensitivity when they have similar sensitivity and communicate confidence. It also says that observers with markedly different sensitivities can perform worse together than the better individual, so the unqualified claim that aggregation improves effectiveness overstates the finding. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [58] Barnett, 2019, Comparative accuracy of diagnosis by collective intelligence of multiple physicians vs individual physicians, JAMA Netw Open source
    • cites [59] Galton, 1907, Vox Populi, Nature source

      In “Vox Populi” (page not specified in our copy), Galton reports that the middlemost estimate of an ox’s weight was within 0.8% of its actual weight, illustrating a benefit of aggregating judgments. The paper describes one weight-judging competition, not general evidence that aggregation improves effectiveness; it also says some participants were experts while others relied on information they picked up or on their own fancies. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [60] Surowiecki, 2005

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    Lack of transparency may lead humans to overreact to AI errors, inferring AI is generally unreliable and incompetent, leading to mistrust and misuse.

    “Third, the lack of transparency may lead humans to overreact to AI errors: if humans cannot understand the cause of an AI error or if they consider it anomalous compared to typical human errors, they may be tempted to infer that the AI is generally unreliable and incompetent, leading to mistrust and misuse of the AI system itself [71 , 72] .”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [71] Madhavan, 2006, Automation failures on tasks easily performed by operators undermine trust in automated aids, Hum Factors source
    • cites [72] Zonca, 2023, Social Influence under uncertainty in interaction with peers, robots and computers, Int J Soc Robot source

      Section 7.5 (page not provided) says people who believe a robot is making mistakes may rapidly lose trust in its competence and stop following its advice. The paper does not link that reaction to a lack of transparency, an inability to explain the error, or errors perceived as anomalous compared with human errors. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation

    Optimal integration of opinions should consider judgments, confidence, and average objective accuracy of agents.

    “Optimal integration of opinions, whether from human or artificial agents, should consider the following factors: their judgments, their confidence in the correctness of their opinions, and the average objective accuracy of the agents in similar situations in the past.”
    Quote in the paperSubstantial claim3 of 5 sources checked
    • cites [44] Behrens, 2008, Associative learning of social value, Nature source

      The paper’s Abstract (page unavailable) says optimal choices in its task combine the likely correct action with the probability that a human confederate’s advice is correct, learned from past outcomes. It does not address an agent’s expressed confidence in a particular opinion or establish the same rule for artificial agents. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [45] Koriat, 2012, When are two heads better than one and why?, Science source
    • cites [46] De Martino, 2017, Social information is integrated into value and confidence judgments according to its reliability, J Neurosci source

      The Behavioral results and Discussion (pages not supplied) support integrating an initial judgment with others’ opinions, weighted by confidence in one’s own judgment and the reliability of the group information. They do not establish that integration should use each agent’s stated confidence or average objective accuracy in similar past situations; the study used the number of Amazon reviews as a measure of group reliability and did not study artificial agents. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [47] Park, 2017, Integration of individual and social information for decision-making in groups of different sizes, PLOS Biol source

      The Discussion (page not available) says participants integrated their own judgments with a group's judgments, weighting them by credibility estimated from their own confidence and group size. The paper does not say they considered agents’ average objective accuracy in similar past situations, and its findings concern human jurors rather than artificial agents. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [48] Sniezek, 2001, Trust, confidence, and expertise in a judge-advisor system, Organ Behav Hum Decis Process source

  • May combine two claims.

    Premisecitation

    Training can reduce interference between tasks, but non-standard procedures can lead to detrimental dual-task interactions and compromise lesion detection abilities.

    “Although training can reduce interference between the tasks, the implementation of non-standard procedures, such as demanding maneuvers, can lead to detrimental dual-task interactions and compromise lesion detection abilities [36] .”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [36] Strobach, 2020, The dual-task practice advantage: empirical evidence and cognitive mechanisms, Psychon Bull Rev source

      The Introduction (page not provided) says dual-task practice can reduce interference, and the Empirical evidence section reports reduced dual-task costs after practice. The paper does not address non-standard procedures or demanding maneuvers compromising lesion detection; its detection examples concern signs or shapes, not lesions. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Human memory for already sampled locations is poor, resulting in less efficient and defective search behavior.

    “However, human memory for already sampled locations proved to be poor, resulting in both less efficient and defective search behavior: less efficient because they may look multiple times at the same location, defective because they may miss a significant proportion of the image [19 , 21 , 22] .”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [19] Lago, 2021, Under-exploration of three-dimensional images leads to search errors for small salient targets, Curr Biol source

      In Results and Discussion (page not available), observers vastly overestimated how much of 3D image stacks they had explored, and extending their search reduced misses for small targets. This supports the claim that incomplete exploration can cause search errors, but the paper does not establish poor memory for already sampled locations or repeated looks at the same location as the cause of inefficiency. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [21] Võ, 2016, You think you know where you looked? You better look again, J Exp Psychol Hum Percept Perform source

      The Abstract and Discussion report poor memory for one's own fixation locations and discuss the possibility that substantial areas of an image go unexamined. They do not establish that this memory failure causes people to revisit the same locations or makes search less efficient; the image-coverage concern is presented as an applied implication, not a measured effect of the experiments. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [22] Williams, 2019, What do we know about volumetric medical image interpretation?: a review of the basic science and medical image perception literatures, Cogn Res Princ Implic source

      In “What are the consequences of limited memory in volumetric image search?” (page unavailable), the review describes poor memory for previously visited locations and suggests that forgetting could lead radiologists to revisit locations or resume searching inaccurately. It presents those effects as possibilities, not proven consequences, and does not establish that poor memory causes searchers to miss a significant proportion of an image. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    When some information is missing, decision-makers must rely on estimates, and humans tend to err in these estimates, often overestimating their chance of being correct, leading to ego bias.

    “Unfortunately, humans tend to err in these estimates, often overestimating their chance of being correct, leading to an implicit, unwarranted preference for their own opinions, an ego bias [52][53][54][55][56][57] .”
    Quote in the paperSubstantial claim2 of 6 sources checked
    • cites [52] Bailey, 2022, A meta-analysis of the weight of advice in decision-making, Curr Psychol source

      The Abstract and Discussion report egocentric discounting: people give their own estimates more weight than others’ advice, even when combining estimates would improve accuracy. The paper does not establish that people overestimate their chance of being correct or that such overconfidence causes the preference for their own opinions. No page numbers are available in our copy. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [53] Bonaccio, 2006, Advice taking and decision-making: an integrative literature review, and implications for the organizational sciences, Organ Behav Hum Decis Process source
    • cites [54] Krueger, 2003, Return of the ego–Self-referent information as a filter for social prediction: comment on Karniol (2003), Psychol Rev source
    • cites [55] Morin, 2021, Social information use and social information waste, Philos Trans R Soc B source

      Section 2 describes egocentric discounting as giving one’s own information more weight than warranted, and Section 3 says self-confidence predicts it. The paper does not establish that people generally overestimate their chance of being correct or that such overestimation causes the bias; it treats why people trust themselves more as an open question. No page number is available for these passages in our copy. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [56] Yaniv, 2004, Receiving other people's advice: influence and benefit, Organ Behav Hum Decis Process source
    • cites [57] Yaniv, 2000, Advice taking in decision making: egocentric discounting and reputation formation, Organ Behav Hum Decis Process source

  • May combine two claims.

    Premisecitation

    The effect of AI advice on experts was smaller but still statistically significant due to the study's high power.

    “The effect on experts was smaller, but it still held significance thanks to the study's high power.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [66] Reverberi, 2022, Experimental evidence of effective human-AI collaboration in medical decision-making, Sci Rep source

      The Results (Effects of expertise, Table 5; page unavailable) show a smaller AI effect on experts than non-experts, while the experts’ diagnostic-accuracy odds ratio is 1.15 with a 95% confidence interval of 1.01–1.30. The Methods (Study design and participants; page unavailable) say the sample size provided adequate power for the main endpoints, but do not establish that high power explains the significant result in experts. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Colonoscopy remains susceptible to errors, prompting significant efforts to reduce them.

    “However, the procedure remains susceptible to errors, prompting significant efforts to reduce them [7 , 8] .”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [7] Kaminski, 2020, Optimizing the quality of colorectal cancer screening worldwide, Gastroenterology source
    • cites [8] Rex, 2015, Quality indicators for colonoscopy, Gastrointest Endosc source

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisecitation

    An evidence-based cognitive model for human-AI interaction would allow designing the most effective human-AI collaboration protocols.

    “First and foremost, establishing an evidence-based cognitive model for human-AI interaction is imperative, as this would allow for designing the most effective human-AI collaboration protocols [68 , 97 , 98] .”
    Quote in the paperNot a substantial claim2 of 3 sources checked
    • cites [68] Cabitza, 2023, Rams, hounds and white boxes: investigating human–AI collaboration protocols in medical diagnosis, Artif Intell Med source

      The Discussion (p. 16) proposes human–AI collaboration protocols as a way to design and evaluate interactions, and the Conclusion (p. 21) reports that testing different protocols identified a more accurate configuration in two diagnostic settings. The paper does not establish an evidence-based cognitive model or show that such a model would enable the design of the most effective protocols generally. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [97] Druce J., Niehaus J., Moody V., Jensen D., Littman M.L. Brittle AI, Causal Confusion, and Bad Mental Models: Challenges and Successes in the XAI Program [Internet]. 2021 [cited 2023 Jul 27];Available from: http://arxiv.org/abs/2106.05506

      Not checked in the free run.

    • cites [98] Sloman, 2015, Causality in thought, Annu Rev Psychol source

      The Introduction discusses models of human causal reasoning, and the Discussion considers limits of causal Bayes nets. Neither addresses human–AI interaction or whether an evidence-based cognitive model would enable more effective collaboration protocols. Searches for human–AI collaboration and cognitive models for its design found no relevant passage in our copy; page numbers were unavailable. (LLM, conf=0.75)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Available AI devices exhibit very high sensitivity and overall performance on par with or better than top human experts.

    “Available AI devices exhibit very high sensitivity and overall performance on par with or better than top human experts [42 , 76] .”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [42] Rex, 2022, Strengths and weaknesses of an artificial intelligence polyp detection program as assessed by a high-detecting endoscopist, Gastroenterology source

      The Testing section reports that the CADe device increased adenoma detection for a high-detecting endoscopist, supporting a performance benefit. But the paper does not establish very high sensitivity or performance on par with or better than top experts across available devices: it assessed one device, reports lesions the device missed, and notes that some endoscopists without AI have exceeded detection rates achieved in AI trials. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [76] Shah, 2023, Effect of computer-aided colonoscopy on adenoma miss rates and polyp detection: a systematic review and meta-analysis, J Gastroenterol Hepatol source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    If both opinions appear equally reliable, a rational decision would integrate them with equal weight, based on statistical principle that aggregating imperfect judgments reduces error.

    “If both opinions appear equally reliable, a rational decision would be to integrate them with equal weight, based on the statistical principle that aggregating imperfect judgments reduces error [49] .”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [49] Larrick, 2006, Intuitions about combining opinions: misappreciation of the averaging principle, Manag Sci source

  • May combine two claims.

    Premisecitation

    Overconfidence may cause fixation and hinder learning and correction, particularly when physicians do not receive systematic feedback.

    “Overconfidence may cause fixation [38] and hinder learning and correction, particularly when physicians do not receive systematic feedback on highly confident yet incorrect judgments [39] .”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [38] Klein, 2007, A data–frame theory of sensemaking source
    • cites [39] Einhorn, 1978, Confidence in judgment: persistence of the illusion of validity, Psychol Rev source

  • May combine two claims.

    Interpretationreasoning

    AI assistance can liberate cognitive resources for other human-exclusive tasks, enhancing overall procedure's efficiency and accuracy.

    “By partially relieving endoscopists from the burden of visual search, AI could liberate cognitive resources for other human-exclusive tasks, such as navigation, framing optimization, and tracking mucosa exploration.This, in turn, can further enhance the overall procedure's efficiency and accuracy.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationnone given

    AI opinion is not infallible but provides a further source of evidence for the final decision of the MD.

    “Importantly, the AI opinion is not infallible but would provide a further source of evidence for the final decision of the MD, as in double-reading or second-opinion scenarios.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Final performance of a human expert assisted by AI depends not only on competence of AI and human but also critically on quality of interaction between the two.

    “While AI holds great promise, it is important to recognize that the final performance of a human expert assisted by AI depends not only on the competence of the AI and the human but also, critically, on the quality of the interaction between the two.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Humans often underestimate susceptibility to errors and tend to be overconfident, reporting higher accuracy estimates than they actually achieve.

    “Unfortunately, humans often underestimate the susceptibility to errors in certain tasks and tend to be overconfident, reporting higher accuracy estimates than they actually achieve.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [37] Dawson, 1993, Hemodynamic assessment in managing the critically iii: is physician confidence warranted?, Med Decis Making source

  • May combine two claims.

    Premisecitation

    About 25% of adenomas are missed in tandem colonoscopies, and 12% of index colonoscopies are false negatives.

    “studies using tandem colonoscopies showed that about 25% of adenomas were missed, and 12% of index colonoscopies (the first performed in the tandem procedure) were false negatives.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [4] Zhao, 2019, Magnitude, risk factors, and factors associated with adenoma miss rate of tandem colonoscopy: a systematic review and meta-analysis, Gastroenterology source
    • cites [5] Adler, 2015, Interval colorectal cancer after colonoscopy: exploring explanations and solutions, Off J Am Coll Gastroenterol ACG source

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    In typical endoscopic conditions, a lesion similar to the surrounding mucosa and 17° away from the fixation point might be easily missed.

    “In typical endoscopic conditions (80 cm large display, 1 m away from the endoscopist), a lesion similar to the surrounding mucosa and 17 ° away from the fixation point might be easily missed.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Recent discussions about the risks of AI may drive more people to close themselves off to AI support systems prejudicially.

    “Recent discussions about the risks of AI may drive more people to close themselves off to AI support systems prejudicially.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    The principles illuminated in the colonoscopy case hold a more general value, extending their relevance to a wide array of medical procedures and beyond.

    “However, the principles illuminated in this case hold a more general value, extending their relevance to a wide array of medical procedures and beyond.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    AI tools demonstrated outstanding standalone detection abilities but might not necessarily provide a measurable advantage in real-world clinical scenarios.

    “Although AI tools demonstrated outstanding standalone detection abilities, they might not necessarily provide a measurable advantage in real-world clinical scenarios.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Cognitive science provides a conceptual framework for comprehending the potentialities and pitfalls of human-AI collaboration, paving the way for its enhancement.

    “Moreover, it provides a conceptual framework for comprehending the potentialities and pitfalls of human-AI collaboration, paving the way for its enhancement.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning

    Cognitive science may provide an important contribution by unraveling the neurocognitive foundations of diagnostic errors and contributing to a theory-informed and evidence-based design of decision support systems.

    “In this context, cognitive science may provide an important contribution, unraveling the neurocognitive foundations of diagnostic errors observed in clinical practice and thus contributing to a theory-informed and evidence-based design of decision support systems.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Processing false positives detected by AI may drain too much attention and time if each triggers a thorough evaluation.

    “Besides, processing false positives detected by AI may drain too much attention and time if each triggers a thorough evaluation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The introduction of AI-based tools to assist MDs during colonoscopy has garnered considerable interest, even though some negative observations have started to emerge.

    “The introduction of AIbased tools to assist MDs during this complex procedure has garnered considerable interest, even though some negative observations also started to emerge [92] .”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [92] Berzin, 2022, Navigating the “Trough of Disillusionment” for CADe polyp detection: what can we learn about negative ai trials and the physician-AI hybrid?, Off J Am Coll Gastroenterol ACG source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Mastering the ability to ignore AI false positives may require appropriate training.

    “However, mastering this remarkable ability may require appropriate training.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    The negative results may be due to strategic misunderstandings, possibly due to inadequate training and poor human-AI fit.

    “The studies suggest that strategic misunderstandings, possibly due to inadequate training and poor human-AI fit, could have contributed to this outcome.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning

    Colonoscopy serves as an insightful application domain and clinical setting due to a specific task situation in which fruitful collaboration between MDs and AI can be easily envisaged and due to the medical community's acute awareness of the importance of addressing diagnostic errors.

    “Colonoscopy serves as an insightful application domain and clinical setting due to a specific task situation in which a fruitful collaboration between MDs and AI can be easily envisaged and due to the acute awareness of the medical community on the importance of addressing diagnostic errors.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    AI devices may benefit MDs differently and should adapt to their attitudes, skills, and profiles.

    “Therefore, it is important to recognize that AI devices may benefit MDs differently and should adapt to their attitudes, skills, and profiles.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Human perceptual, attentional, and navigation skills or ability to detect patterns are unequal, as is level of expertise in a procedure.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    Such understanding will help identify situations where AI may be prone to errors or exceptionally dependable and improve the human-AI fit.

    “Such understanding will help identify situations where AI may be prone to errors or exceptionally dependable and improve the human-AI fit.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning

    By identifying multiple bottlenecks in perception, attention, and decision-making that contribute to diagnostic errors, cognitive science provides crucial insights for designing effective interventions, leading to optimal integration of AI systems into clinical practice.

    “By identifying the multiple bottlenecks in perception, attention, and decision-making that contribute to diagnostic errors, cognitive science provides crucial insights for designing effective interventions, thus leading to the optimal integration of AI systems into clinical practice.”
    Quote in the paperNot a substantial claim
33 of 33

  • The paper relies on this without support.

    Human visual search is highly effective for single salient feature targets.

    “Human visual search is highly effective when seeking a target stimulus that differs from an irrelevant background based on a single, perceptually salient feature.”

  • The paper relies on this without support.

    AI tools can search for multiple complex visual targets without attentional bottlenecks and without performance degradation in the periphery of the visual field.

    “These tools can search for multiple complex visual targets without succumbing to attentional bottlenecks and without experiencing performance degradation in the periphery of the visual field.”

  • The paper relies on this without support.

    Modern AI learns and reaches conclusions in a different, non-human way.

    “By contrast, modern AI learns and reaches conclusions in a different, non-human way.”

  • The authors state this as an assumption.

    AI assistance can be conceptualized as a human expert who evaluates similar evidence and provides advice but leaves the final decision to the collaborator.

    “To better understand this, we can conceptualize AI assistance as that of a human expert who evaluates similar evidence and provides advice but leaves the final decision to the collaborator.”

  • The paper relies on this without support.

    Human experts remain the ultimate decision-makers.

    “Effective AI assistance should target the primary causes of human errors and foster effective collaborative decision-making with human experts who remain the ultimate decisionmakers.”

  • The paper relies on this without support.

    Existing insights on human-AI interaction remain preliminary, and a comprehensive understanding of the cognitive bottlenecks leading to performance disparities is lacking.

    “While existing insights provide valuable perspectives, they remain preliminary, and a comprehensive understanding of the cognitive bottlenecks ( Table 1 ) leading to performance disparities is lacking.”

  • The paper relies on this without support.

    Human perceptual, attentional, and navigation skills or ability to detect patterns are unequal, as is level of expertise in a procedure.

    “Our perceptual, attentional, and navigation skills or ability to detect patterns are unequal, as is our level of expertise in a procedure.”

  • The authors state this as an assumption.

    In the figure example, the endoscopist is fixating at the center of the lumen while scanning the mucosa.

    “In this example, we assume that the endoscopist is fixating at the center of the lumen (fixation cross) while scanning the mucosa.”
8 of 8

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Introzzi, L., Zonca, J., Cabitza, F., Cherubini, P., & Reverberi, C.
Affiliations
  • University of Milano-Bicocca (IT)
  • IRCCS Ospedale Galeazzi - Sant'Ambrogio (IT)
  • University of Pavia (IT)
Corresponding author
Reverberi, Carlo
ORCID
Venue
Digestive and Liver Disease
Publisher
Elsevier BV
Published
2024-07
Volume
56
Issue
7
Pages
1131-1139
ISSN
1590-8658 (ISSN-L)
DOI
10.1016/j.dld.2023.10.018
OpenAlex
W4388445128
Semantic Scholar
56f93cbad326ec1bb2520ea7cfc73cebade06cbe
Cited by
28
References
122
Citations by year
2026: 8 · 2025: 18 · 2024: 2
FWCI
0.33
Type
journal article
Language
en
Open access
Licence
Locations
  • Digestive and Liver Disease (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • BOA (University of Milano-Bicocca) (submittedVersion, cc-by)
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Health Sciences › Medicine › Oncology › Colorectal Cancer Screening and Detection
  • Health Sciences › Medicine › Family Practice › Clinical Reasoning and Diagnostic Skills
Keywords
human-AI collaboration, colonoscopy, diagnostic errors, AI-assisted decision-making, human-AI integration, diagnostic accuracy
MeSH terms
Clinical Decision-Making, Artificial Intelligence, Colonoscopy, Diagnostic Errors, Humans

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-11-03.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-11-03.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review