Human–artificial intelligence adaptation in colonoscopy: beyond deskilling and upskilling

10.1055/a-2858-7084 Computer-aided polyp detection (CADe) in colonoscopy has been enthusiastically heralded as one of the most significant emerging endoscopic technologies, supported by an impressive body of evidence from multiple randomized controlled trials (RCTs) [ 1 ]. However, the transition from impressive trial efficacy to widespread clinical adoption has proven more challenging. Real-world “pragmatic” observational studies have failed to consistently replicate the detection gains seen in RCTs [ 2 ]. Furthermore, a recent study suggests that the introduction of CADe could lead to inadvertent “deskilling” or erosion of endoscopists’ independent detection capabilities, raising concerns among endoscopists [ 3 ] [ 4 ]. Conversely, some proponents argue that CADe exposure may promote upskilling, producing a learning effect that persists during AI-unassisted procedures. To address this, Pedersen et al. conducted a pragmatic, prospective, multicenter clinical trial specifically designed to evaluate CADe-associated deskilling or learning effects [ 5 ]. PDR-5 (the proportion of colonoscopies with detection of at least one polyp ≥5 mm) was measured across three consecutive phases (before, during, and after CADe exposure) for 13 endoscopists performing 5013 colonoscopies. The study demonstrated a transient increase in PDR-5 among inexperienced endoscopists only during CADe use. Importantly, there was no upskilling or deskilling observed in unassisted procedures following CADe exposure. „The key question may no longer be whether AI detects more lesions, but what kind of endoscopist it ultimately creates.“ The differential effect of CADe according to operator experience is an intriguing finding. Are less experienced operators simply more receptive to CADe prompts, or are seasoned endoscopists dismissing them because of algorithm aversion? The common explanation of a ceiling effect is not entirely convincing: if experienced endoscopists were already operating near the limits of detection, inexperienced operators should not have consistently achieved higher PDR-5 rates across every study phase. Perhaps the critical question may no longer be whether CADe works in the real world, but for whom, under what circumstances, and through which cognitive mechanisms. The absence of either upskilling or deskilling contrasts with findings from two recent studies, although apparent discrepancies can be explained by differences in study designs. Pedersen et al. suggest that exclusive short-term CADe exposure improves procedural performance more readily than it reshapes fundamental perceptual skills [ 5 ]. In contrast, Budzyń et al. evaluated endoscopists performing non-AI procedures while also working within a persistently AI-assisted environment, reporting evidence of deskilling or dependency among experienced operators [ 3 ]. Rather than true skill erosion, however, this may reflect transient cognitive offloading, a reversible adaptation to AI support rather than permanent deskilling. Longer-term data offer a different perspective. In a 3-year observational study, Okumura et al. demonstrated preserved unassisted detection rates following CADe implementation, alongside accelerated learning among high detectors [ 6 ]. Together, these findings suggest CADe does not exert a uniform effect. Instead, its influence likely depends on baseline expertise, attentional strategies, and perhaps individual cognitive traits. Viewed through this lens, the literature may be less contradictory than it first appears; these studies may simply capture different stages of human–AI adaptation. There are of course limitations to the current study, which the authors acknowledge. PDR-5 is not a robustly evaluated quality indicator. Moreover, inexperienced endoscopists may have overestimated polyp size, artificially inflating PDR-5 rates. CADe predominantly improves the detection of diminutive (≤5 mm) polyps; therefore PDR-5 may not fully capture the effects of CADe-associated skill modification. The pragmatic study design also resulted in fewer inexperienced endoscopists performing procedures in phase 3, potentially reducing statistical power. Despite this, the authors should be commended, especially for their three-phase study design, which allowed endoscopists to serve as their own controls, whilst capturing data from diverse clinical settings including an academic center, general hospital, and a high-volume outpatient practice. In conclusion, the study by Pedersen et al. compels us to re-evaluate how we interact with AI as it becomes embedded into routine endoscopic practice. Many have argued that medical algorithms are the new drugs, requiring comparable post-marketing surveillance and pharmacovigilance. Yet this analogy may be flawed. Unlike drugs, CADe systems actively shape human decision making, involving complex interactions between operator, algorithm, and clinical environment. AI is not merely an intervention; it is a participant in the cognitive process. The unique behavior-modifying properties challenge traditional frameworks for evaluating endoscopic innovations and demand new dedicated study designs informed by human–computer interaction and implementation science [ 7 ] [ 8 ]. The key question may no longer be whether AI detects more lesions, but what kind of endoscopist it ultimately creates. Publication History Article published online: 16 July 2026 © 2026. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany Refers to: Learning and deskilling effects of artificial intelligence in colonoscopy among endoscopists with different levels of experience: a pragmatic, prospective trial Endoscopy eFirst DOI: 10.1055/a-2858-7084