How study design differences explain conflicting AI colonoscopy results

In August 2025, The Lancet Gastroenterology & Hepatology published a finding that travelled the world: at four Polish centres, endoscopists who had spent months working with AI-assisted detection found fewer adenomas when performing a colonoscopy without it. The rate fell from 28.4% to 22.4%. Four months later, a Norwegian prospective trial with 5,013 colonoscopies and three phases found exactly nothing: neither loss nor gain of skill when the tool was withdrawn. And a third study, over three years, found sustained improvement. This article sorts out a scientific controversy that is live right now, and explains what separates the three designs.

The 30-second summary: the seven figures that matter

Each row identifies the study and its publication year. Observe that the initial three contradict each other, although all three were executed with high quality.

What you will find here

• The Polish study that raised the alarm and how it is built.

• The three-phase prospective trial that did not find the effect.

• What has been measured outside endoscopy: radiology and pathology.

• The uncomfortable detail about which lesions are detected more and less.

The study that raised the alarm

Krzysztof Budzyń, Marcin Romańczyk alongside a broad team released a retrospective observational study in The Lancet Gastroenterology & Hepatology during 2025, which took place across four Polish endoscopy centres involved in the ACCEPT trial. Polyp detection tools were adopted by those facilities at the close of 2021, prompting the researchers to evaluate colonoscopy standards delivered without AI throughout the preceding three months against the subsequent three months following that deployment.

The sample: 1,443 patients, 795 before and 648 after, with a median age of 61. The primary outcome was the adenoma detection rate in standard, unassisted colonoscopy.

The result: detection fell from 28.4% (226 of 795) to 22.4% (145 of 648), an absolute difference of −6.0 percentage points (95% CI: −10.5 to −1.6; p = 0.0089). In multivariable analysis, prior AI exposure remained an independent factor associated with that fall, with an odds ratio of 0.69.

What that design cannot separate

It is a retrospective observational study comparing two three-month periods. Anything that changed at those centres between late 2021 and early 2022 — staffing, workload, the kind of patient referred — enters the comparison alongside the AI. The authors phrase it cautiously: continuous exposure “might” reduce detection, “suggesting” a negative effect.

The trial that did not find the effect

Tor Anders Pedersen, Yuichi Mori along with their colleagues released a pragmatic, prospective, multicentre study in Endoscopy during late 2025 that was specifically formulated to resolve this inquiry. A total of 5,013 colonoscopies were performed across three sequential stages—pre-AI, during AI deployment, and post-withdrawal—by thirteen endoscopists, comprising seven novices and six veterans who had each completed upwards of a thousand procedures. Consequently, every clinician served as their own control.

The authors’ conclusion is clear: detection by inexperienced endoscopists increased while they were using the tool, and in unassisted colonoscopy neither improvement nor loss of skill was observed after the exposure period.

This layout surpasses the Polish alternative for this particular inquiry, owing to its prospective nature and the fact that every practitioner serves as their own benchmark. Furthermore, it involves a smaller cohort—thirteen compared to the participants across four facilities—and relies on an indirect metric, namely the fraction of colonoscopies revealing at least one polyp measuring 5 mm or greater, which lacks the established standing as a quality benchmark that adenoma detection enjoys.

What has been measured outside endoscopy

Pierre Heudel and colleagues published a scoping review on deskilling in medicine in ESMO Real World Data and Digital Oncology in 2026. Their general conclusion: evidence of clinical deskilling is scarce, but consistent across specialties. Three specific figures they record:

• In radiology, a controlled study with 27 breast imaging radiologists found that erroneous AI prompts increased false-positive recalls by up to 12%, even among experienced readers.

• In computational pathology, in experimental tasks under time pressure, more than 30% of participants reversed a correct initial diagnosis when exposed to an incorrect system suggestion.

• In cytology, the United Kingdom’s shift to primary human papillomavirus screening cut case volumes by 80% to 85% and consolidated laboratories from 45 to 8 centres, with direct consequences for training capacity.

That last point introduces a distinction worth keeping. A professional losing their edge by leaning on the machine is not the same as a system that stops generating the cases people learn on. The second problem is structural and cannot be fixed with individual training.

The uncomfortable detail: what is detected more and less

An observational investigation spanning three years in Japan, encompassing upwards of 4,700 paired colonoscopies, revealed zero decline following the adoption of assisted detection, while additionally demonstrating expedited learning curves for endoscopists who already possessed strong detection capabilities. This appeared to decisively resolve the controversy.

Darrien Henry and James East published a letter in Endoscopy in 2026 interpreting those identical statistics through a different lens. Within that research, detection rates for adenomas measuring 6 to 9 mm, 10 mm and above, those featuring high grade dysplasia, along with invasive cancer cases, proved markedly superior among the unassisted cohort. Furthermore, the incidence of invasive cancer was more than twice as high when the tool was absent compared to when it was utilized (p = 0.001).

Their alternative interpretation is harsh and well argued: assisted detection may distract the endoscopist towards trivial lesions — adenomas of 1 to 5 mm and hyperplastic polyps — at the expense of advanced lesions, which are the ones that matter. The consistency of the effect direction across three different categories of advanced lesion, with very low p values, makes chance an unlikely explanation.

Omer Ahmad proposed in 2026, also in Endoscopy, the reading that best reconciles the three studies: they may not contradict one another, but rather be capturing different stages of human-machine adaptation, and what the Polish study measured may be transient, reversible cognitive offloading rather than permanent erosion of skill.

What to do with this, in practice

Ordered by the strength of the evidence behind it:

• Do not view the discussion as completely resolved on either side. Three rigorous analyses point to three distinct outcomes, and their divergent methodologies account for the variance.

• Keep periods of unassisted practice. It is the recommendation that recurs in the endoscopy literature, precisely because of what is not known.

• Measure what kind of lesion you detect, rather than simply counting them. The specialty’s most rigorous review revealed that a higher volume of invasive cancer was uncovered without the software than with its assistance.

• Watch for erroneous prompts. Within breast radiology, they elevated false-positive recalls by as much as 12%, even among seasoned practitioners.

• Think about training, not only about the already-trained professional. The British cytology case shows the problem may be that the cases people learn on disappear.

Frequently asked questions

Do doctors who use artificial intelligence lose skill?

It remains unconfirmed. A Polish observational study featured in The Lancet Gastroenterology & Hepatology during 2025 revealed that unassisted adenoma detection rates dropped from 28.4% to 22.4% following the integration of AI, representing a 6.0-point decrease (p = 0.0089), whereas a subsequent prospective trial encompassing 5,013 colonoscopies detected neither decline nor improvement after the technology was removed.

Why do two good studies reach opposite conclusions?

Because of the design. The Polish one is retrospective and compares two three-month periods, so any change at the centre enters the comparison; the Norwegian one is prospective, three-phase, and each endoscopist acts as their own control. A hypothesis published in Endoscopy holds that both may be capturing different stages of adaptation to the tool.

Who does it impact the most?

The evidence points to different effects by level. In the prospective trial, inexperienced endoscopists improved detection while using the AI, from 31.9% to 39.5% (OR 1.43), and returned to their baseline when it was withdrawn; among the experienced there was no significant effect in any of the three phases.

Does it happen outside endoscopy too?

There are signals in other specialties. According to the 2026 review, a controlled study with 27 breast imaging radiologists found erroneous AI prompts increased false-positive recalls by up to 12% even among experienced readers, and in computational pathology more than 30% of participants reversed a correct diagnosis when faced with an incorrect suggestion under time pressure.

Does finding a greater number of polyps translate to stopping more cancer?

Not necessarily, and that is the most uncomfortable point. Reanalysing the data from a three-year study, a letter published in Endoscopy in 2026 showed that detection of adenomas 10 mm or larger, of high grade dysplasia and of invasive cancer was significantly higher without the tool, with an invasive cancer rate more than double (p = 0.001).

What we do not know

• Whether the effect is reversible. The transient cognitive offloading hypothesis remains plausible, yet nobody has measured what happens after withdrawing the tool for an extended period.

• What happens to people trained with the tool from the start. The available studies do not include absolute beginners, the group where skill retention matters most.

• Whether it affects mortality. None of this work measures post-colonoscopy colorectal cancer or survival.

• How long the exposure needs to be. One study, which observes an impact, spans three months; another, showing no effect, evaluates varying durations; finally, the research noting enhancement lasts three years.

References

1. Budzyń, K., Romańczyk, M. et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology, 2025. DOI: 10.1016/s2468-1253(25)00133-5

2. Pedersen, T. A., Mori, Y. et al. Learning and deskilling effects of artificial intelligence in colonoscopy among endoscopists with different levels of experience: a pragmatic, prospective trial. Endoscopy, 2025. DOI: 10.1055/a-2858-7084

3. Heudel, P., Crochet, H. et al. Artificial intelligence in medicine: a scoping review of the risk of deskilling and loss of expertise among physicians. ESMO Real World Data and Digital Oncology, 2026. DOI: 10.1016/j.esmorw.2026.100693

4. Henry, D. and East, J. Artificial intelligence-supported polyp detection (CADe) at colonoscopy reduces the detection of high grade dysplasia and invasive cancer. Endoscopy, 2026. DOI: 10.1055/a-2749-8345

5. Ahmad, O. Human-artificial intelligence adaptation in colonoscopy: beyond deskilling and upskilling. Endoscopy, 2026. DOI: 10.1055/a-2896-9791

This material is purely informational and should not be considered medical guidance. The statistics mentioned stem from research featuring methodologies, fields, and timeframes that cannot be directly contrasted, and they outline an ongoing scientific discussion at the moment of publication.