Answers are grounded in this paper and cite their sources.
Dear Editor,Artificial intelligence (AI) is rapidly transforming the practice of interventional pulmonology (IP) [1‒3]. Emerging applications – including AI-assisted lung nodule detection and radiomics-based risk stratification, robotic and computer-aided bronchoscopic navigation, AI-enhanced endobronchial ultrasound image interpretation, machine learning-supported rapid on-site evaluation and cytopathology assessment, and decision-support tools for ultrasound-guided pleural procedures – are increasingly reshaping the field by enhancing diagnostic accuracy and procedural precision [4‒6]. Although not yet universally adopted, these technologies are widely anticipated to become progressively embedded in everyday procedural workflows and to redefine the procedural landscape in the coming years.As this trajectory unfolds, the IP community has a unique and time-sensitive opportunity. While the potential benefits for diagnostic accuracy and workflow efficiency are increasingly recognized, a critical and underappreciated risk deserves attention: the possibility that sustained AI exposure may gradually erode the procedural and perceptual skills that define expert-level interventional practice.The phenomenon of AI-induced deskilling – defined as the degradation of previously acquired clinical competencies due to reduced independent practice consequent to automation – was systematically examined in a recent landmark mixed-method review by Natali and colleagues [7]. This work identified several interdependent mechanisms through which AI may impair clinical skills: automation bias (the tendency to over-trust algorithmic outputs), cognitive offloading (the progressive delegation of perceptual and reasoning tasks to the machine), and upskilling inhibition (the reduction in opportunities to develop expert-level skills, particularly among trainees). Equally important, deskilling is not merely an individual cognitive phenomenon – it also carries organizational and systemic dimensions, threatening team situational awareness, the mentorship pipeline, and institutional resilience when AI systems fail or produce erroneous outputs [7].The most instructive empirical signal currently available comes from an adjacent procedural specialty. Budzyń and colleagues [8] recently reported results of a multicentre observational study across four endoscopy centers participating in the ACCEPT trial, demonstrating that sustained exposure to AI-assisted colonoscopy was associated with a significant decline in adenoma detection rate during subsequent non-AI-assisted procedures – from 28.4% before AI implementation to 22.4% after (OR 0.69, 95% CI 0.53–0.89; p = 0.0089). AI exposure emerged as an independent predictor of performance decline on multivariable analysis.This finding – the first prospective empirical demonstration of procedural deskilling attributable to AI exposure in a clinical setting – attracted immediate regulatory and bioethical attention, including scrutiny under the European Union Artificial Intelligence Act [9]. It serves as a timely reminder that benefits and risks of AI integration in procedural medicine can coexist and that the latter may manifest before anticipated.Although no analogous data exist for IP – and it would be premature to assume that findings from colonoscopy translate directly – there are strong a priori reasons to believe that IP may be particularly vulnerable to similar deskilling dynamics. Several structural features converge to heighten this concern. First, IP procedures are fundamentally dependent on continuous real-time integration of perceptual cues, spatial reasoning, and contextual clinical judgment: precisely the cognitive and psychomotor domains most susceptible to atrophy under conditions of habitual AI reliance. Second, many of these procedures are performed at moderate volumes even in specialized centers, limiting the natural density of deliberate practice that might otherwise buffer against skill decay. Third, learning curves for IP skills are steep and protracted; trainees who develop procedural competence in an AI-saturated environment may never fully acquire independent perceptual benchmarks, experiencing what Natali et al. [7] term upskilling inhibition rather than deskilling per se. Finally, AI systems in interventional settings have finite sensitivity and specificity, and their failure modes are not always transparent. A clinician whose independent interpretive skills have atrophied is precisely the least equipped to recognize when an AI output is erroneous.The conceptual distinction between AI as an augmentation tool and AI as a cognitive substitute is not merely academic – it has direct implications for how AI should be introduced, monitored, and governed in procedural training and practice. AI that prompts clinicians to reason, challenges interpretations, and preserves the human in an active decision-making role is fundamentally different from AI that simply provides an answer to be accepted or overridden. The former has genuine potential to enhance clinical care; the latter risks producing a workforce that is technically capable under AI-assisted conditions but progressively less resilient in their absence.Looking ahead, the IP community should proactively address this risk before it becomes empirically measurable through adverse outcomes. From a research standpoint, prospective longitudinal studies should monitor procedural performance metrics stratified by degree of AI exposure, with dedicated AI-off assessment phases analogous to the Budzyń study design [8]. From an educational standpoint, training curricula should explicitly incorporate structured AI-off sessions, ensuring trainees regularly perform procedures requiring fully autonomous perceptual and clinical reasoning. Simulation-based training, validated across surgical specialties as a robust strategy to mitigate skill decay, should be leveraged in IP to maintain procedural competence independently of live AI-assisted case volume. From a governance standpoint, international scientific societies and credentialing bodies should consider recommending minimum non-AI-assisted procedure volumes as part of ongoing competency maintenance, drawing on the precedent established by aviation regulators, who mandate minimum manual flying hours to counteract automation-induced skill fade [10].This letter does not argue against AI adoption in IP – the potential of these technologies to improve diagnostic accuracy, reduce operator dependency, and broaden procedural access is both real and worth pursuing. The point, rather, is that AI integration into procedural medicine is not a zero-risk proposition and that its risks – including deskilling – are modifiable if addressed early.The field should adopt the same degree of foresight in managing the risks of AI integration as it does in evaluating its benefits. The colonoscopy data provide proof of concept that procedural deskilling is not merely a theoretical risk but a measurable clinical phenomenon. IP should not wait for analogous data before implementing the educational and governance safeguards that may prevent it. In the age of AI, preserving the irreplaceable elements of human procedural expertise is not a resistance to progress – it is a prerequisite for safe, high-quality care.The author declares no conflicts of interest.The author received no specific funding for this work.The author is solely responsible for the conception, drafting, and final approval of the manuscript.
No takes yet. Share an insight, caveat, or question.
Guido Marchi (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: