What we corrected on 19 October 2026

Was: “After the AI was removed it was 36.3% (173 of 476) against the 31.9% baseline — an adjusted odds ratio of 1.03, on 0.79 to 1.34, which is no change.”

Now: “After the AI was removed it was 36.3% (173 of 476) against the 31.9% baseline — an unadjusted odds ratio of 1.03, on 0.79 to 1.34, not a statistically significant change.”

Was: “The system detected polyps faster than people did — a median 1.16 seconds against 2.97 — but the people using it were no faster than the people working without it (2.90 seconds, indistinguishable from 2.97).”

Now: “The system detected polyps faster than people did — a median 1.16 seconds against 2.97 — but the people using it were no faster than the people working without it (2.90 seconds, not statistically different from 2.97).”

Was: “When the authors tested selectively withholding AI outputs that were likely to mislead, automation bias fell by 41.7%.”

Now: “When the authors simulated selectively withholding AI outputs that were likely to mislead, recalculating their own data on the assumption that clinicians would fall back on their unassisted diagnoses, automation bias was estimated to fall by 41.7%.”

Was: “Access to it raised their confidence by 11.7 percentage points while about half its answers were faulty — and their confidence did not fall as they met more wrong answers.”

Now: “Access to it raised their confidence by 11.7 percentage points while about half its answers were faulty — and their confidence did not fall significantly as they met more wrong answers.”

Was: “The same is true of the research highlight in Nature Reviews Gastroenterology & Hepatology of 5 September 2025 and of Decker J, “When AI Reads the Easy Cases”, Rad Insights, 15 May 2026 — which pairs the colonoscopy and mammography results and is the closest published thing to this page.”

Now: “The research highlight in Nature Reviews Gastroenterology & Hepatology of 5 September 2025 calls it a multicentre, observational, retrospective study; we hold only the publisher’s preview, which says nothing more about its limitations. Decker J, “When AI Reads the Easy Cases”, Rad Insights, 15 May 2026, calls it a multicentre observational study but mentions none of these others: the near-doubling of procedures after the screening programme was suspended, the missing withdrawal times, the absence of blinding, or the exclusion of the first 100 procedures. It pairs the colonoscopy and mammography results and is the closest published thing to this page that we found.”

Was: “Source of the three phases — before CADe, during CADe, after CADe removal — the thirteen endoscopists (seven inexperienced, six experienced) and 5,013 colonoscopies, the primary endpoint of detection of at least one polyp of 5 mm or more, the rise among inexperienced endoscopists from 31.9% (216/678) to 39.5% (257/651) at an odds ratio of 1.43 on 1.11–1.84, the 36.3% (173/476) after removal at an adjusted odds ratio of 1.03 on 0.79–1.34, the absence of any significant effect among the experienced, and the quoted conclusion.”

Now: “Source of the three phases — before CADe, during CADe, after CADe removal — the thirteen endoscopists (seven inexperienced, six experienced) and 5,013 colonoscopies, the primary endpoint of detection of at least one polyp of 5 mm or more, the rise among inexperienced endoscopists from 31.9% (216/678) to 39.5% (257/651) at an odds ratio of 1.43 on 1.11–1.84, the 36.3% (173/476) after removal at an odds ratio of 1.03 on 0.79–1.34, the absence of any significant effect among the experienced, and the quoted conclusion.”

Was: “Gerlich's 2025 study is the source of the figure most often quoted, a correlation of −0.68 between AI use and critical thinking.”

Now: “Gerlich's 2025 study reports a correlation of −0.68 between AI tool use and critical thinking.”

Was: “Unassisted solve rates fell, at effect sizes of −0.42, −0.19 and −0.42, and people gave up sooner.”

Now: “Unassisted solve rates fell, at effect sizes of −0.45, −0.20 and −0.42, and people gave up more often — significantly in the first and third experiments, not in the second.”

Later updated to the authors’ revised version.

Was: “Separately, three pre-registered experiments with 1,372 participants across 9,593 trials randomised whether an AI assistant's answer was right or wrong: participants consulted it on more than half of trials and, having consulted it, followed it on 79.8% of the trials where it was wrong.”

Now: “Separately, three pre-registered experiments with 1,372 participants across 9,593 trials randomised whether an AI assistant's answer was right or wrong: participants consulted it on more than half of trials and, in the first experiment, having consulted it, followed it on 79.8% of the trials where it was wrong.”

Was: “Wharton School Research Paper, February 2026 — Source of the three pre-registered experiments, 1,372 participants and 9,593 trials, the 79.8% of chat-engaged trials on which participants followed faulty advice, and the confidence rise of 11.7 percentage points that did not decline as faulty trials accumulated.”

Now: “Wharton School Research Paper, February 2026 — Source of the three pre-registered experiments, 1,372 participants and 9,593 trials, and, from the first experiment, the 79.8% of chat-engaged trials on which participants followed faulty advice and the confidence rise of 11.7 percentage points that did not significantly decline as faulty trials accumulated.”

Was: “Hesselmann's 2020 replication analysed 89 participants of 117 recruited and did not find the effect: the data were about five times more likely under the null than under the alternative, and about sixteen times more likely when tested against the original's specific predicted pattern.”

Now: “Hesselmann's 2020 replication analysed 89 participants of 117 tested and did not find the effect: the data were about five times more likely under the null than under the alternative, and, in an exploratory analysis, about sixteen times more likely when tested against the original's specific predicted pattern.”

Was: “The authors list eleven limitations of their own, and the largest is not subtle: the total number of procedures in the second period nearly doubled, from 795 to 1,382 — while the non-AI procedures actually measured fell from 795 to 648 — because Poland suspended its colonoscopy-based screening programme on 1 January 2022 and the centres used the freed capacity to draw patients off the waiting list.”

Now: “The authors list eleven limitations of their own, and the one we think matters most is not subtle: the total number of procedures in the second period nearly doubled, from 795 to 1,382 — while the non-AI procedures actually measured fell from 795 to 648 — primarily because Poland suspended its colonoscopy-based screening programme on 1 January 2022 and the centres used the freed capacity to take more patients from the waiting list.”

Was: “Three randomised experiments — 1,222 people randomised, 1,060 analysed — had participants work through practice problems with or without AI, then removed the AI without warning and gave them three more.”

Now: “Three randomised experiments — 1,222 people randomised, 1,058 analysed — had participants work through practice problems with or without AI, then removed the AI and gave them three more.”

Later updated to the authors’ revised version.

Was: “arXiv 2604.04721 — Source of the three randomised experiments, 1,222 randomised and 1,060 analysed, the effect sizes of −0.42, −0.19 and −0.42 on unassisted solve rates, and the design in which the AI is withdrawn without warning before three test problems.”

Now: “arXiv 2604.04721 — Source of the three randomised experiments, 1,222 randomised and 1,058 analysed, the effect sizes of −0.45, −0.20 and −0.42 on unassisted solve rates, and the design in which the AI is withdrawn before three test problems.”

Later updated to the authors’ revised version.

Was: “And when the AI's errors exceeded an acceptability threshold, they overrode it.”

Now: “And when the AI's errors were serious, beyond about three points on the global score, their own errors largely stopped following it; just past the acceptability threshold of two points, they still tracked it.”

Was: “The journal says interpretation is unchanged and we have no basis to disagree; we note it because a correction to the adjusted estimate for the thing a study is about is worth a reader knowing, whichever way it went.”

Now: “The journal says interpretation is unchanged and we have no basis to disagree; we note it because a correction to a study’s adjusted model is worth a reader knowing, whichever way it went.”

Was: “Time pressure did not make that happen more often; it made it worse, with heavier reliance on the system's negative calls and a measurable fall in performance.”

Now: “Time pressure did not make that happen more often, but it appeared to make it worse, with heavier reliance on the system's negative calls and a fall in performance.”

Was: “It shows no significant effect in any phase — but that includes no benefit while the AI was running, and the authors suggest why: experienced endoscopists may also be less enthusiastic about a software tool intended to improve their performance, and two endoscopists ran it throughout while the rest switched it on only during withdrawal.”

Now: “It shows no significant effect in any phase — but that includes no benefit while the AI was running, and the authors suggest why: experienced endoscopists may also be less enthusiastic about a software tool intended to improve their performance . They also report that two endoscopists had CADe active throughout the entire examination, including intubation, while the rest switched it on only on withdrawal.”

Was: “Never-skilling is a trainee who never develops the skill, because the tool was there from the beginning.”

Now: “Never-skilling is a trainee who fails to develop the skill in the first place, because the AI did the thinking while the skill was being formed.”

Was: “A critical appraisal for the American College of Gastroenterology laid out the design problems for gastroenterologists and proposed the gaze mechanism.”

Now: “A critical appraisal for the American College of Gastroenterology laid out the design problems for gastroenterologists and suggested, heavily hedged, that changes in visual gaze could be a mechanism.”

Was: “Time carried both on 13 August 2025, quoting Professor Osmani on the rise in procedure volume and the fatigue it could cause, and Professor Tucker on automation bias not being peculiar to AI, and calling the study observational throughout.”

Now: “Time carried both on 13 August 2025, reporting Professor Osmani‘s point about the rise in procedure volume and the fatigue it could cause, quoting Professor Tucker on automation bias not being peculiar to AI, and calling it an observational study, though its headline said using AI made doctors worse.”

Was: “Carries Professor Osmani’s volume-and-fatigue caveat by name, quotes Professor Tucker on automation bias not being peculiar to AI, and calls the study observational throughout.”

Now: “Carries Professor Osmani’s volume-and-fatigue caveat by name, quotes Professor Tucker on automation bias not being peculiar to AI, and calls it an observational study, though its headline says using AI made doctors worse.”

Was: “What did not travel is why it rose: not one piece of coverage we read named the suspension of the Polish screening programme as the reason, or the absence of withdrawal-time data, or the eleven limitations as a body.”

Now: “What did not travel is why it rose: none of the news coverage we read named the suspension of the Polish screening programme as the reason, the absence of withdrawal-time data or the eleven limitations as a body, and STAT’s article continues behind a paywall we could not read; only the American College of Gastroenterology’s appraisal noted the missing withdrawal times.”

Was: “A radiologist writing for trainees paired the colonoscopy result with the mammography experiment — the closest thing to this page in general-audience writing — and reported the limitations of neither.”

Now: “A radiologist writing for trainees paired the colonoscopy result with the mammography experiment — the closest thing to this page that we found — and, beyond calling the colonoscopy study observational, reported none of the limitations of either.”

Was: “It is cross-sectional, and the author does not claim causation.”

Now: “Its survey is cross-sectional, so it cannot show causation — though the paper writes of “the negative impact of AI usage on critical thinking”, and itself calls for experiments that “could offer causal evidence”.”

Was: “The author does not claim causation.”

Now: “The design cannot show causation, and the author calls for experiments that “could offer causal evidence”, but the paper writes of “the negative impact of AI usage on critical thinking”.”

Was: “The Polish endoscopists were still working with the AI most days and being measured on the days they were not.”

Now: “The Polish endoscopists were still working with the AI on the days allocated to it, about half their lists, and being measured on the days they were not.”

Was: “Professor Venet Osmani of Queen Mary University of London named the volume problem exactly — the number of colonoscopies performed nearly doubled after the AI tool was introduced, going from 795 to 1382 — and argued that three months is too short a window for genuine skill loss in experienced clinicians.”

Now: “Professor Venet Osmani of Queen Mary University of London named the volume problem exactly — the number of colonoscopies performed nearly doubled after the AI tool was introduced, going from 795 to 1382 — and questioned whether a true loss of skill could happen in three months, a very short period, especially for a clinician with over 27 years of experience (the study’s endoscopists averaged 27.6 years, from 8 to 39).”

Was: “They also point at Troya and colleagues — the eye-tracking study in the section above — as the mechanism that would explain the Polish result.”

Now: “They also point at Troya and colleagues — the eye-tracking study in the section above — as a mechanism that could explain the Polish result.”

Was: “Separately, four learning scientists audited the studies behind the next-most-cited synthesis and found that of 19 comparisons, four met three basic criteria: a well-defined treatment, an adequately described control, and a valid learning measure.”

Now: “Separately, four learning scientists audited a subset of the studies behind another synthesis and found that of 19 comparisons, four met three basic criteria: a well-defined treatment, an adequately described control, and a valid learning measure.”

Was: “Editorial Retraction Note, Humanities and Social Sciences Communications, 22 April 2026 — The retraction of the most-cited meta-analysis finding that ChatGPT improves learning.”

Now: “Societies 2025;15(9):252 — The notice replacing the paper’s Table 4, which had duplicated Table 3, and carrying the author’s statement that the scientific conclusions are unaffected. Editorial Retraction Note, Humanities and Social Sciences Communications, 22 April 2026 — The retraction of the meta-analysis finding that ChatGPT improves learning.”

Was: “Nobody's skill decayed; there was simply far less of the work on which the skill is built.”

Now: “Nobody measured an individual's skill decaying; there was simply far less of the work on which the skill is built — which, the College warns, may itself reduce diagnostic expertise.”

Was: “This one is — specifically on the question of what happens after the tool is withdrawn, since it found a real gain while the tool was running — with the limits its own authors put on it.”

Now: “This one is — specifically on the question of what happens after the tool is withdrawn, since among its inexperienced endoscopists it found a real gain while the tool was running — with the limits its own authors put on it.”

Was: “If AI exposure changes what an endoscopist can do unassisted, something has to have changed in how they look.”

Now: “If AI exposure changes what an endoscopist can do unassisted, one plausible place for the change is in how they look.”

Was: “Nobody has tested whether those are the same thing, and this page does not claim they are.”

Now: “We have found no study that tests whether those are the same thing, and this page does not claim they are.”

Was: “It is a study about what happens inside a single reading, and it is the clearest measurement in this entire field of the thing people actually describe when they say AI has changed how they work: you take what it gives you.”

Now: “It is a study about what happens inside a single reading, and it measures directly what, in our reading, people describe when they say AI has changed how they work: you take what it gives you.”

Was: “The evidence a general reader has been offered is much weaker than the evidence they have not.”

Now: “And the rest of us In our reading, the evidence a general reader is usually offered is much weaker than the evidence that has mostly not reached them with its qualifications intact.”

Was: “The belief that AI is degrading thinking does not usually travel on any of the above.”

Now: “In our reading, the belief that AI is degrading thinking does not usually rest on any of the above.”

Was: “It travels on three studies, and all three are weaker than their reputations.”

Now: “Three studies often cited for it are each weaker than they are usually taken to be.”

Was: “That uniformity is what common-method variance looks like.”

Now: “That uniformity is what common-method variance would produce, though it does not prove it is present.”

Was: “The strongest evidence that AI use degrades a real skill, in real work, comes from an endoscopy suite — measured not after the tool was taken away but on the days it was not used, during a period when it was in use on the others.”

Now: “The strongest evidence we found that AI use degrades a real skill, in real work, comes from an endoscopy suite — measured not after the tool was taken away but on the days it was not used, during a period when it was in use on the others.”

Was: “The ones that would make a reader hold the number loosely did not.”

Now: “The ones that would make a reader hold the number loosely did not, in the coverage we read.”

Was: “What did not travel is the specific.”

Now: “What mostly did not travel is the specific.”

Was: “And the further the finding got from the specialty, the less arrived with it.”

Now: “And the pieces we read that were furthest from the specialty, in business and computing outlets, carried the least of it.”

Was: “That is the same design error this page is about, in a magazine read by more people than the journal — and a draft of this section said nothing false had been printed here.”

Now: “That is the same design error this page is about, in a general business magazine — and a draft of this section said nothing false had been printed here.”

Was: “None of this makes the review's substance wrong, and we are relying on it.”

Now: “None of this undermines the parts of the review we rely on — its count of included studies and its description of the evidence as scarce and fragmented.”

Was: “It makes the point that carrying a finding accurately across a boundary is difficult enough that the careful people fail at it too.”

Now: “It shows that carrying a finding accurately across a boundary is difficult enough that a specialist review, whose own body gets the design right, can still fail at it.”

Was: “It settles the confounding that the volume doubling introduced, and it is the single study that would move this from an association to a causal finding.”

Now: “It would address the confounding that the volume doubling introduced, and it is the design that would most directly test whether the association is causal.”

Was: “Eye tracking during real colonoscopy, linked to what was actually found, would move the mechanism from plausible to shown.”

Now: “Eye tracking during real colonoscopy, linked to what was actually found, would test the proposed mechanism directly rather than in video clips.”

Was: “It does not say whether any clinician, department or reader should use these tools, and nothing here bears on whether AI-assisted colonoscopy is good for patients — the randomised evidence that AI raises detection during the procedure is not in dispute and is not what this page examined.”

Now: “It does not say whether any clinician, department or reader should use these tools, and nothing here bears on whether AI-assisted colonoscopy is good for patients — the randomised evidence that AI raises detection during the procedure is not something this page disputes and is not what it examined.”

Was: “The studies being quoted in general coverage are not the strong ones, in either direction.”

Now: “The studies quoted in the general coverage we read are not the strong ones.”

Was: “Both of those are true, and neither is what a reader takes from a review date on a page.”

Now: “Both of those are true, and in our judgement neither is what a reader takes from a review date on a page.”

Was: “The clearest evidence about what happens when the machine is wrong comes from a reading room.”

Now: “Among the studies we read, the clearest evidence about what happens when the machine is wrong comes from a reading room.”

Was: “The case mix moved between phases.”

Now: “Among the inexperienced endoscopists, the case mix moved between phases.”

Was: “A larger prospective study found no fall in endoscopists measured after the AI was withdrawn, most clearly among the inexperienced, on a different and weaker endpoint.”

Now: “A prospective study with more than twice the procedures, though fewer endoscopists, found no significant fall in endoscopists measured after the AI was withdrawn, most clearly among the inexperienced, on a different and weaker endpoint.”

Was: “The first ten cases carried correct suggestions, which established the system's credibility.”

Now: “The first ten cases carried correct suggestions, to establish the system's credibility.”

Was: “That is Dratsch's forty-point swing, and the pathologists who abandoned correct judgments, and the endoscopists who saw more polyps that were not there.”

Now: “Its first step — taking the machine’s errors in the moment — is what Dratsch's swing of 37 to 60 points, the pathologists who abandoned correct judgments, and the endoscopists who saw more polyps that were not there all measured; none of them tested whether the errors were learned.”

Was: “Sparrow, Liu and Wegner's 2011 finding that people offload memory to the internet is the origin of the modern version of this worry.”

Now: “Sparrow, Liu and Wegner's much-cited 2011 finding that people offload memory to the internet is an influential source of the modern version of this worry.”

Was: “The most-cited meta-analysis finding that ChatGPT improves learning — cited several hundred times — its own journal's counter read 402 at one point on 30 August 2026 and 399 an hour later, which is what live metrics do, and other counts of the same paper run from 262 to over 500 — was retracted by its journal on 22 April 2026.”

Now: “A widely cited meta-analysis finding that ChatGPT improves learning — cited several hundred times — was retracted by its journal on 22 April 2026.”

Was: “It did not reach a general reader.”

Now: “It did not reach a general reader in any coverage we read.”

Was: “The 2023 study the description otherwise fits is not in the review's reference list at all.”

Now: “The 2023 study whose year, reader count and finding the description matches — 27 radiologists, though reading 50 mammograms rather than 720 — is not in the review's reference list at all.”

Was: “Source of the proposal that visual gaze is the mechanism, citing Troya and colleagues.”

Now: “Source of the suggestion that the effect on visual gaze could potentially be a mechanism, citing Troya and colleagues.”

Was: “The best evidence anyone has is real, and almost nobody outside medicine has met it with its qualifications intact: in four Polish endoscopy centres, the rate at which experienced doctors found precancerous growths without AI fell from 28.4% to 22.4% after their units adopted it.”

Now: “In our reading the strongest evidence is real, and in the coverage we read it has mostly not reached readers outside medicine with its qualifications intact: in four Polish endoscopy centres, the rate at which experienced doctors found precancerous growths without AI fell from 28.4% to 22.4% after their units adopted it.”

Was: “That study is retrospective, confounded, and its authors list eleven limitations.”

Now: “That study is retrospective, its authors call it vulnerable to confounding, and they list eleven limitations.”

Was: “Three named academics said so publicly the day it appeared, and some of that did travel.”

Now: “Two of three named academics raised its limits publicly the day it appeared, and some of that did travel.”

Was: “What did not is the part that tells you how loosely to hold the number.”

Now: “What mostly did not, in the coverage we read, is the part that tells you how loosely to hold the number.”

Was: “Source of the 89 participants analysed of 117 recruited, and of the Bayes factors of 5.07 against the alternative and 16.43 against the original’s specific predicted pattern.”

Now: “Source of the 89 participants analysed of 117 tested, and of the Bayes factors of 5.07 against the alternative and 16.43 against the original’s specific predicted pattern.”

Was: “Insights into Imaging 2025;16:23 — Source of the eight radiologists in training, the 150 chest X-rays across three conditions, the intraclass correlation rising from 0.665 without AI to 0.813 with it integrated, the 70% of cases in which residents requested the AI, and their overriding it when its errors exceeded an acceptability threshold.”

Now: “Insights into Imaging 2025;16:23 — Source of the eight radiologists in training, the 150 chest X-rays across three conditions, the intraclass correlation rising from 0.665 without AI to 0.813 with it integrated, the 70% of cases in which residents requested the AI, and the finding that once its errors were serious, beyond about three points on the global score, their own errors largely stopped following it, while just past the acceptability threshold of two points they still tracked it.”

Was: “The December 2024 preprint of the same work is on SSRN and its conclusion is worded more strongly than the published one. Expert reaction Science Media Centre — expert reaction to the colonoscopy study, 12 August 2025 — Source of the three named comments published on the day of release: Professor Venet Osmani of Queen Mary University of London on the near-doubling of procedures and on three months being too short a window; Professor Allan Tucker of Brunel on the single AI system and the need for randomised crossover trials; Dr Catherine Menon of the University of Hertfordshire on the underlying concern.”

Now: “The December 2024 preprint of the same work is on SSRN and its conclusion is worded more strongly than the published one. Expert reaction Science Media Centre — expert reaction to the colonoscopy study, 12 August 2025 — Source of the three named comments published on the day of release: Professor Venet Osmani of Queen Mary University of London on the near-doubling of procedures and on whether a true loss of skill could happen in three months; Professor Allan Tucker of Brunel on the single AI system and the need for randomised crossover trials; Dr Catherine Menon of the University of Hertfordshire on the underlying concern.”

Was: “The effect is real, statistically clear, and modest.”

Now: “The effect is statistically significant and, in our reading, modest, and it was measured on video clips.”

Was: “So, within the twelve studies that review includes: not one study, and not a settled literature either.”

Now: “So, within the twelve studies that review includes: more than one study, and not a settled literature either.”

Was: “The evidence usually offered is weak.”

Now: “In our reading, the evidence usually offered for it is weak.”

Was: “This page uses their hedges, not firmer ones.”

Now: “On consequences this page uses their hedges, not firmer ones; that the narrowing they measured is the mechanism behind the Polish result is Zhou’s and Pedersen and colleagues’ proposal, not a finding.”

Was: “We expected to find that this evidence had reached the public without its qualifications, and the first thing we looked for was whether that was true.”

Now: “We expected to find that this evidence had reached the public without its qualifications, and we first gathered coverage that fitted that expectation; a later, deliberate search for the contrary case tested whether it was true.”

Was: “We caught it by opening the article again rather than trusting our note of it.”

Now: “A fact-check run flagged it, and we confirmed it by opening the article again rather than trusting our note of it.”

Was: “We read the review's full text at source on 29 August 2026, and noticed only because we had read both underlying papers that evening.”

Now: “We read the review's full text at source on 29 August 2026, and noticed only because we had read the main results of both underlying papers that evening.”

Was: “We check the sources on this page against their journals' notices, and record when we last did.”

Now: “We check the sources on this page that carry a DOI or PubMed identifier for corrections and retractions in Europe PMC and Crossref, and record when we last did; most of the others cannot be looked up this way, and the lookup failed for three.”

Was: “If a general-audience account carries this evidence with its limitations attached, we will say so and link to it.”

Now: “Whether a general-audience account carries this evidence with its limitations attached is one of the questions we watch; we last looked on 30 August 2026.”

Was: “The full watch list, with the searches that would answer each question, is kept in the repository that produces this page, alongside a record of every check — including the ones that found nothing.”

Now: “The full watch list, with the searches that would answer each question, is kept in the repository that produces this page, alongside a record of the checks we have run — including the ones that found nothing — although three fact-check runs on 30 August left no report.”

Was: “Every figure above comes from the study that produced it, read directly.”

Now: “Every study figure above comes from the study that produced it, read directly, except the 70% cytology estimate, which is Public Health England’s as the Royal College of Pathologists reports it.”

Was: “Where we could not read something, we say so rather than passing on somebody else's account of it.”

Now: “Where we could not read something, we say so rather than passing on somebody else's account of it — with one exception now corrected: the Gastroenterology editorial, which we know only by its title and authors.”

Was: “What we watch is narrower than what could change — we can see a registry record move and a paper cite another paper, and we cannot see most of what happens in a field — and nothing obliges anyone to look before you arrive.”

Now: “What we watch is narrower than what could change — we can see a paper cite another paper, though no registry record is yet captured for this page, and we cannot see most of what happens in a field — and nothing obliges anyone to look before you arrive.”

Was: “The watch list below is real and it runs; it is our instrument, not a promise to you.”

Now: “The watch list above is real, and its citation and correction sweeps run weekly; it is our instrument, not a promise to you.”

Was: “Where we read a study through somebody else’s account of it rather than at source, this says so.”

Now: “Where we read a study through somebody else’s account of it rather than at source, this says so, with one exception: the 70% cytology estimate is Public Health England’s, as the Royal College of Pathologists reports it.”

Was: “Read in the accepted manuscript held by the University of Warwick repository; the publisher’s own site refused every request.”

Now: “Read in the accepted manuscript held by the University of Warwick repository, which gives the P value as 0.009; the published version is paywalled and we have not read it.”

Was: “This page relies on the review for that count and for nothing else; each paper named above was read at source.”

Now: “This page relies on the review for that count and for its authors’ characterisation of the evidence; the studies named above were read at source, except Rebolj and colleagues, which we could not reach, and the Weinberg and Bretthauer editorial, which we know only by its record.”

Was: “Read at the publisher.”

Now: “Read in the published version, as deposited in PubMed Central.”

Was: “Everything this page says about that work is a description of what the replications set out to reproduce, taken from the replication papers themselves.”

Now: “Everything this page says about that work is a description of what the replications set out to reproduce, taken from Hesselmann’s replication paper.”

Was: “No pre-registration statement appears in any version.”

Now: “No pre-registration statement appears in version 2, 4 or 6, the three we read.”

Was: “Three things on this page fall into that category: a commentary published alongside the colonoscopy study, which we could not open; the 2011 memory study, which is paywalled; and one meta-analysis whose scope we describe as its auditors reported it rather than as we read it.”

Now: “Among the things on this page that fall into that category are a commentary published alongside the colonoscopy study, and the later correspondence on it, which we could not open; the 2011 memory study, which is paywalled; and one meta-analysis whose scope we describe as its auditors reported it rather than as we read it.”

Was: “The abstract was read at Europe PMC on 30 August 2026 and the full paper the same evening, after the operator obtained it; the publisher's own page returned 404 to every URL form we tried.”

Now: “The abstract was read at Europe PMC on 30 August 2026 and the full paper the same day, after the operator obtained it.”

Was: “Its abstract describes the colonoscopy study as “a multicenter randomized trial”; the study is retrospective and observational, as the review’s own body says.”

Now: “Its abstract describes the colonoscopy study as “a multicenter randomized trial”; the study is observational, as the review’s own body says, and retrospective, as the study itself states.”

Was: “arXiv 2601.00856 — Their closing sentence — that there is no strong evidence the tool use negatively affects performance itself — is a verdict on the paper they audited, not on the field.”

Now: “arXiv 2601.00856 — Their sentence that there is no strong evidence the LLM use negatively affects performance itself is a verdict on the paper they audited, not on the field.”

Was: “Societies 2025;15(1):6 — Source of the 666 valid responses of 669, the recruitment through social media in the United Kingdom, the cross-sectional design, and the three correlations of −0.68, +0.72 and −0.75.”

Now: “Societies 2025;15(1):6 — Source of the 666 valid responses of 669, the recruitment through social media in the United Kingdom, the mixed-method design — a single 23-question questionnaire and 50 semi-structured interviews — and the three correlations of −0.68, +0.72 and −0.75.”

Was: “Dratsch and colleagues showed 27 radiologists a set of mammograms alongside AI-generated BI-RADS category suggestions.”

Now: “Dratsch and colleagues had 27 radiologists read a set of mammograms alongside BI-RADS category suggestions from a purported AI system.”

Was: “Nobody was randomised.”

Now: “The before-and-after comparison was not randomised, although after installation the days on which AI was switched on were allocated at random by date.”

Was: “The authors' call for randomised crossover trials did travel, but only inside the specialty: Professor Tucker made it at the Science Media Centre on the day of publication, and the American College of Gastroenterology's appraisal asked for prospective before-and-after trials.”

Now: “The authors' call for randomised crossover trials did travel, but not far: Professor Allan Tucker, a computer scientist at Brunel, relayed it at the Science Media Centre on the day of publication, and the American College of Gastroenterology's appraisal asked for prospective trials comparing colonoscopy quality before and after AI exposure.”

Was: “It is a retrospective observational study; the review's body says so.”

Now: “It is a retrospective observational study; the review's own body calls it observational.”

Was: “The colonoscopy authors asked for it, and Professor Tucker asked for it independently on the day of publication.”

Now: “The colonoscopy authors asked for it, and Professor Tucker repeated their call at the Science Media Centre on the day of publication.”

Was: “But one of the two findings that was affected is the exposure variable itself, AI introduction, in the adjusted model.”

Now: “But one of the two findings that was affected is the AI-assisted category, colonoscopies performed with AI after its introduction, in the supplementary adjusted model covering all 2,177 colonoscopies.”

Was: “The paper’s title calls it a scoping review; its methods section describes a narrative review.”

Now: “The paper’s title calls it a scoping review, and so does its methods section; the methods paragraph of its abstract describes a narrative review.”

Was: “The Norwegian endoscopists had it taken away entirely — all but two of the thirteen, who reported running it throughout.”

Now: “The Norwegian endoscopists had it taken away entirely: all thirteen did a third phase after CADe was removed.”

Was: “One team measures the narrowing with the tool present and calls it efficiency; another measures the outcome with the tool absent and worries about what was not looked at.”

Now: “One team measures the narrowing with the tool present and calls it a more efficient search, while noting it could let cancers the AI misses go unseen; another measures the outcome with the tool absent and worries about what was not looked at.”

Was: “Also the source of the authors' own reconciliation with the Polish study, of their limitations — PDR-5 rather than ADR, the three-and-a-half-year time effect, the case-mix shift, possible underpowering, one device — and of the endoscopist-level split.”

Now: “Also the source of the authors' own reconciliation with the Polish study, of their limitations — PDR-5 rather than ADR, the time effect, the case-mix shift, possible underpowering, one device — and of the endoscopist-level split.”

Was: “The sentence under it was not.”

Now: “A sentence further down was not.”

Was: “Neither reaches the horizon people actually ask about, which is years.”

Now: “Neither follows endoscopists through years of AI use before measuring them without it.”

Was: “Where it was, 7% of judgments the pathologists had initially got right were overturned by it.”

Now: “Of the 560 estimates they made with its aid, about 7% were judgments the pathologists had initially got right and that its erroneous advice overturned.”

Was: “Those are compatible, and the second study's authors say so.”

Now: “Those are compatible, and the second study's authors explain why.”

Was: “The correction notice states that the scientific conclusions are unaffected, and nothing on this page turns on the corrected table.”

Now: “The correction notice records the author’s statement that the scientific conclusions are unaffected, and nothing on this page turns on the corrected table.”

Was: “The scoping review gives 80–85% and 45 laboratories, citing a 2022 cytopathology paper we could not reach at the publisher, on PubMed or in PMC; we print the College’s figures because we could read them.”

Now: “The scoping review gives 80–85% and 45 laboratories, citing in its body an NHS Digital annual report and a UK National Screening Committee evidence review and in its Table 1 a 2022 cytopathology paper we could not reach at the publisher, on PubMed or in PMC; we print the College’s figures because we could read them.”

Was: “A comment on it by four researchers at Vienna and Dresden calculates that roughly 159 participants would have been needed for adequate power, notes that some reported figures rest on two to four essays per group, and observes that the paper reports 55 completers while analysing 54, and that the criteria for dropping the one remain unclear.”

Now: “A comment on it by four researchers at Vienna and Dresden, itself a preprint, calculates that roughly 159 participants would have been needed for adequate power, notes that some reported figures rest on two to four essays per group, and observes that the paper reports 55 completers while analysing 54, and that the criteria for dropping the one remain unclear.”

Was: “Greengard S, “The AI Deskilling Paradox”, Communications of the ACM, 7 November 2025, carries none of this study’s limitations but does carry caveats of its own, on governance and on deskilling having good outcomes as well as bad ones; we had described it as carrying none, and that was wrong.”

Now: “Greengard S, “The AI Deskilling Paradox”, Communications of the ACM, 7 November 2025, carries none of this study’s limitations but does carry caveats of its own, on governance and on deskilling having good outcomes as well as bad ones; we had described it as carrying no caveats at all, and that was wrong.”

Was: “Source of the power calculation of roughly 159 participants, the observation that some reported figures rest on two to four essays per group, the discrepancy between 55 completers and 54 analysed — on which the paper itself gives a reason the comment does not address, and this page prints the paper’s — the definition of a “significant connection” as a relative measure and the consequence they draw from it, and their assessment that there is no strong evidence that the tool use negatively affects performance itself.”

Now: “Source of the power calculation of roughly 159 participants, the observation that some reported figures rest on two to four essays per group, the discrepancy between 55 completers and 54 analysed — on which the paper itself gives a reason the comment does not address, and this page prints the paper’s — the definition of a “significant connection” as a relative measure and the consequence they draw from it, and their assessment that there is no strong evidence that the LLM use negatively affects performance itself.”

Was: “That is a drop of 6.0 percentage points, 95% confidence interval −10.5 to −1.6 percentage points — both ends below zero — at P = 0.0089.”

Now: “That is a drop of 6.0 percentage points, 95% confidence interval −10.5 to −1.6 percentage points — both ends below zero — at P = 0.009.”

Was: “Lancet Gastroenterol Hepatol 2025;10(10):896–903 — Source of the 28.4% (226/795) to 22.4% (145/648) fall, the −6.0 percentage-point difference on a 95% confidence interval of −10.5 to −1.6 at P = 0.0089, the nineteen endoscopists with more than 2,000 procedures each and a mean 27.6 years since graduation over a range of 8 to 39, and all eleven limitations including the rise from 795 procedures to 1,382 after the Polish screening programme was suspended on 1 January 2022, the absence of withdrawal-time data, the absence of blinding and the exclusion of the first 100 procedures after AI was introduced.”

Now: “Lancet Gastroenterol Hepatol 2025;10(10):896–903 — Source of the 28.4% (226/795) to 22.4% (145/648) fall, the −6.0 percentage-point difference on a 95% confidence interval of −10.5 to −1.6 at P = 0.009, the nineteen endoscopists with more than 2,000 procedures each and a mean 27.6 years since graduation over a range of 8 to 39, and all eleven limitations including the rise from 795 procedures to 1,382 after the Polish screening programme was suspended on 1 January 2022, the absence of withdrawal-time data, the absence of blinding and the exclusion of the first 100 procedures after AI was introduced.”

Was: “The mechanism was measured in 2022, by people who then said what worried them about it.”

Now: “A possible mechanism was measured in a study published in 2022, by people who then said what worried them about it.”

Was: “The same radiologists, the same task, and a forty-point swing.”

Now: “The same radiologists, the same task, and a swing of between 37 and 60 points.”

Was: “Primary Royal College of Pathologists.”

Now: “Primary Saeed-Chesterman D (Royal College of Pathologists).”

Was: “Fifty-four participants completed its first three sessions; the fourth session, which carries the claim about lasting effects, had eighteen, split across two crossover arms.”

Now: “Fifty-four participants are reported for its first three sessions (fifty-five completed them); the fourth session, which carries the claim about lasting effects, had eighteen, split across two crossover arms.”

Was: “ESMO Real World Data and Digital Oncology 2026;12:100693 — Source of the count of included work — twelve studies from 373 records screened — and of the authors’ own characterisation of the evidence as scarce, limited and fragmented.”

Now: “ESMO Real World Data and Digital Oncology 2026;12:100693 — Source of the count of included work — twelve studies from 373 records identified — and of the authors’ own characterisation of the evidence as scarce, limited and fragmented.”

Was: “The proof we hold prints “Volume 11 ■ Issue C” in its running footer, above “Available online xxx”; the publisher’s registered metadata for the version of record gives volume 12 and no issue, which is what we cite.”

Now: “The copy we hold prints “Volume 12 ■ Issue C ■ 2026” in its running footer, and the publisher’s registered metadata for the version of record gives volume 12, which is what we cite.”

Was: “it says it reports 54 to ensure data distribution (as participants were assigned in three groups)”

Now: “it says it reports 54 participants to ensure data distribution, because participants were assigned in three groups”

Was: “Healthgrades, 19 August 2025 — Calls it “a retrospective observational study” and carries an outside expert’s judgement that “the study design, population limits, operator factors, and short-term focus restrict broad applicability, and the deskilling concern remains hypothetical rather than proven”.”

Now: “Healthgrades, 19 August 2025 — Calls it a “retrospective observational study” and carries an outside expert’s judgement that “the study design, population limits, operator factors, and short-term focus restrict broad applicability, and the deskilling concern remains hypothetical rather than proven”.”

Was: “Their December 2024 preprint had said reduced, and a negative impact on endoscopist performance; the version that came through peer review says may reduce, and behaviour.”

Now: “Their December 2024 preprint had said reduced, and a negative impact on endoscopist performance; the version that came through peer review says may reduce, and behavior.”

Was: “Extended and peer reviewed as Rosbach E, Ammeling J, Ganz J, Bertram CA, Conrad T, Riener A, Aubreville M, “Stuck on Suggestions”, accepted at the Journal of Machine Learning for Biomedical Imaging, 2026 (arXiv:2603.11821) — Source of the 28 trained pathology experts, the 7% automation-bias rate — initially correct evaluations overturned by erroneous AI advice — and the finding that time pressure did not raise the frequency of automation bias but appeared to increase its severity, through heavier reliance on the system’s negative consultations and a subsequent fall in performance.”

Now: “Extended and peer reviewed as Rosbach E, Ammeling J, Ganz J, Bertram CA, Conrad T, Riener A, Aubreville M, “Stuck on Suggestions”, Machine Learning for Biomedical Imaging, published 13 March 2026 (arXiv:2603.11821) — Source of the 28 trained pathology experts, the 7% automation-bias rate — initially correct evaluations overturned by erroneous AI advice — and the finding that time pressure did not raise the frequency of automation bias but appeared to increase its severity, through heavier reliance on the system’s negative consultations and a subsequent fall in performance.”

Was: “That is causal, and it held in all three experiments — which is internal replication by one team in one preprint, not confirmation by anybody else.”

Now: “That is causal, and the fall in solve rates held in all three experiments — which is internal replication by one team in one paper, not confirmation by anybody else.”

Was: “It is also about fifteen minutes long, and it is a preprint with no pre-registration statement.”

Now: “Its second experiment took about fifteen minutes, and the paper, published at COLM 2026, carries no pre-registration statement.”

Was: “Three of the studies this page relies on are preprints or working papers, and we have said so each time.”

Now: “Two of the studies this page relies on are a preprint and a working paper, as is the comment on one of them, and we say so in the text where each appears; a third, by Liu and colleagues, has been published as a COLM 2026 conference paper.”

Was: “Preprint Liu, Christian, Dumbalska, Bakker & Dubey.”

Now: “Conference paper Liu, Christian, Dumbalska, Bakker & Dubey.”

Was: “The same narrowing has been measured in breast imaging, and read there as a gain.”

Now: “The same narrowing has been measured in breast imaging, and read there mainly as a gain, though its authors note it could also raise the risk of missing cancers the AI does not detect.”

Was: “One meta-analysis in this literature has already been retracted after nine and a half months of public criticism, and a scoping review in it mis-cites a paper.”

Now: “One meta-analysis in this literature has already been retracted after nine and a half months of public criticism, and the scoping review that gathers this literature for doctors misdescribes the colonoscopy study and misattributes figures in its key-studies table.”

Was: “Also the source of and the quoted sentence about differences measured during treatment reflecting the availability of the treatment rather than durable learning.”

Now: “Also the source of the quoted sentence about differences measured during treatment likely reflecting the availability of the treatment rather than durable learning.”

Was: “The two counts differ and we cannot say why: the College's report and the review's source disagree about the starting number, and NHS Digital's own annual report says 48 consolidated to 8 without mentioning any earlier closures.”

Now: “The two counts differ and we cannot say why: the College's report gives 48 accredited labs reduced to 8 centralised labs as of 2019, while the review's source, which we could not read, starts from 45.”

Was: “And it has already been applied to this exact literature, by someone with standing to do it.”

Now: “And the frame has already been named in this literature, in the title of an editorial by Weinberg and Bretthauer, which we have not been able to read.”

Was: “What this page does that it does not is carry the distinction out of the specialty to a reader deciding something about their own work — and that, rather than the frame, is the part we can claim.”

Now: “What this page tries to add to an editorial written for the specialty is to carry the distinction out of it to a reader deciding something about their own work — and that, rather than the frame, is the part we can claim.”

Was: “So the specialty is working this frame over its own evidence, in its own journals, and has been since the week the null result was published.”

Now: “So, to judge by that editorial’s title (we have not read its text), the specialty has been putting a version of this frame to its own evidence, in its own journal, since the week the null result was published.”

Was: “A research highlight in Nature Reviews Gastroenterology & Hepatology labelled the study observational and retrospective, as it should, and then reported the figures with no limitations and no quotes.”

Now: “A research highlight in Nature Reviews Gastroenterology & Hepatology labelled the study observational and retrospective, as it should, and then, in the opening we could read, reported the figures with no limitations and no quotes.”

Was: “The review’s body cites this paper correctly by its arXiv number, though under the wrong author names and a wrong title; its Table 1 attaches the same 28 pathologists and the same 7% to a different reference — Bellahsen-Harrar and colleagues, PLOS ONE 2025;20(8):e0323270, which studied eight pathologists, had no time-pressure condition and reports no such figure.”

Now: “The review’s body cites this paper correctly by its arXiv number, though under the wrong author names and a wrong title; its Table 1 attaches the same 28 pathologists and the same 7% to a different reference — Bellahsen-Harrar and colleagues, PLOS ONE 2025;20(8):e0323270.”

Was: “Source of the date and of the editor’s stated ground.”

Now: “Not held; the date and the editor’s stated ground are as the audit’s front matter reports them.”

Was: “Checked against version 4.”

Now: “Read in version 4 (5 August 2026) and version 6 (13 October 2026); the figures here are version 6’s.”

The entry of 12 September 2026 says: “The entry of 12 September below said “What we watch is narrower than what could change — we can see a trial registry move and a paper cite another paper, and we cannot see most of what happens in a field — and nothing obliges anyone to look before you arrive.” That was wrong: no trial-registry record is captured for this page, so we cannot see one move; the entry is left as written, and the correction is listed with the others.”

Correction: no trial-registry record is captured for this page, so we cannot see one move. The body of the page now says so: “What we watch is narrower than what could change — we can see a paper cite another paper, though no registry record is yet captured for this page, and we cannot see most of what happens in a field — and nothing obliges anyone to look before you arrive.”

Was: “The best evidence that using AI degrades a real skill comes from endoscopy suites and reading rooms, not from the studies everyone quotes.”

Now: “In our reading, the strongest evidence that using AI degrades a real skill comes from endoscopy suites and reading rooms, not from the studies everyone quotes.”

Was: “titled Upskilling, Deskilling, or Never Skilling: Who Benefits From Artificial Intelligence in Colonoscopy?”

Now: “titled Upskilling, Deskilling, or Never Skilling: Who Benefits From Artificial Intelligence in Colonoscopy?”

Unchanged: “Among them, and each read here at source rather than through the review’s summary of it:”

When we wrote that, Wang and colleagues’ paper was not among the documents we held. We now hold it, and it carries the two figures that follow the sentence as stated; the third is corrected under the first heading on this page.