When AI does the thinking, who stays sharp?

The colonoscopy study that changed the clinical AI conversation

In 2025, a multicenter observational study published in The Lancet Gastroenterology & Hepatology provided some alarming insights with respect to the clinical AI community. Researchers tracked 19 experienced endoscopists — each with more than 2,000 lifetime colonoscopies — across four Polish endoscopy centers. After these physicians used AI-assisted polyp detection tools, their adenoma detection rate in unassisted procedures dropped from 28.4% to 22.4%, a 6-point absolute decline.¹

The AI had worked exactly as designed. It improved detection while in use. But something else happened in the background: when the assist was removed, the physicians performed worse than before they’d ever used it.

The study authors attributed the drop to “the natural human tendency to over-rely” on decision support systems — what one co-author described as “the Google Maps effect.” When you always use GPS, navigating by instinct becomes harder.¹

This is the deskilling question that matters most in clinical AI. Not whether AI-assisted documentation frees up cognitive bandwidth (it does). Not whether ambient scribes reduce late-night charting hours (they do). The deeper question is: what happens to a physician’s diagnostic judgment when an algorithm is always present, always confident, and always faster?

The deskilling phenomenon occurs across industries, with growing evidence in healthcare

Deskilling isn’t a new anxiety. The aviation industry has raised concerns regarding how increased autopilot usage may result in declines to situational awareness and the ability to re-engage at complex, time-sensitive moments. Similarly, in medicine, AI-induced skill erosion can create specific out-of-the-loop vulnerabilities in physical examination, differential diagnosis, clinical judgment, and physician-patient communication.²

Perhaps most counterintuitively, AI can mask the problem entirely. Improved patient outcomes may no longer reflect stable physician competence. They may instead reflect increasing dependence on AI for diagnostic reasoning, creating what researchers call “the illusion of competence.”³

A 2025 qualitative study interviewing clinicians and AI experts found that practicing physicians today generally don’t report deskilling from current AI tools, partly because their clinical foundations predate AI. Their concern is directed forward: toward the next generation of physicians who will train with AI from day one.³ 

That concern has a name too. Researchers call it “never-skilling,” and argue it poses a greater long-term threat to medical education than deskilling itself.⁴ It occurs when trainees rely on automation so early in their development that they never acquire foundational clinical reasoning in the first place.⁴ 

Simulation research on physician-AI decision dynamics reinforces this: in training environments, junior clinicians may disproportionately benefit from AI assistance and face disproportionate risk of dependency. Deskilling is not inevitable, but it is more likely when AI confidence signals are not calibrated to preserve independent judgment.⁵ 

Where EHRs fit — and why it matters 

Healthcare AI doesn’t operate in isolation. The platforms that ingest and act on clinical data, notably EHRs, play an important role in both shaping how insights are generated and ensuring they’re applied responsibly in care delivery. 

In a physician survey, concerns about deskilling broke down into three roughly equal categories: reduced vigilance and increased automation bias (22%), deskilling of new physicians (22%), and erosion of clinical judgment and empathy (22%). One surgical oncologist described it plainly: “I noticed I stopped second-guessing and my diagnostic muscles dulled.”⁶ 

Documentation AI is part of this picture. Ambient scribes and AI-assisted charting do shift cognitive load: by automating repetitive administrative tasks, they can free physicians to refocus on complex decision-making, procedural excellence, and patient interaction.⁴ Early adoption programs at UCLA Health⁷ and Mass General Brigham⁸ have shown ambient scribes can help reduce documentation time. Some of those proof points may explain why adoption and frequency of comfort with ambient scribes seems to be growing. The 2026 Physician Sentiment Survey (PSS) found 38% increased usage of AI for generating clinical documentation. 

The risk isn’t that AI writes the note. The risk is when AI writes and finalizes the clinical conclusion and the physician is no longer the decision-maker.

The human-in-the-loop imperative 

AI should augment clinical reasoning: improving diagnostic accuracy, supporting triage, and freeing clinician time for more complex tasks. What distinguishes augmentation from replacement, in practice, comes down to where in the decision chain the human is positioned, and how actively they’re required to engage. 

A 2025 randomized controlled trial in Communications Medicine found that physicians were willing to modify their clinical decisions based on GPT-4 assistance, and that AI assistance improved accuracy meaningfully across patient groups.⁹ Crucially, it did so without introducing or exacerbating demographic bias in triage decisions. This is the case for AI as genuine decision support: not replacing physician judgment but sharpening it with data the physician uses actively. 

The colonoscopy study points to the opposite failure mode: passive reliance that atrophies the underlying skill. Researchers conclude that safeguarding clinical expertise should be considered a central component of AI safety, and that longitudinal monitoring, adaptive training curricula, and clear regulatory frameworks are necessary to prevent skill erosion from becoming systemic.¹ 

What responsible AI design looks like in practice 

For clinicians, administrators, and the technology partners who build and deploy AI in clinical settings, a few principles hold: 

• Keep the physician the decision-maker. AI recommendations in areas such as diagnostic suggestions, risk flags, or documentation summaries, should be surfaced as inputs to assist with physician reasoning, not replace independent decision-making.
• Audit AI-assisted workflows for performance drift. Physicians who want to preserve clinical judgment can treat AI as a sophisticated second opinion: interrogating its recommendations, asking whether they align with guidelines and the specific patient’s context, and resisting the pull to accept without evaluating.
• Maintain unassisted practice. The colonoscopy study’s finding suggests that continuous AI exposure without periods of independent practice can erode the skill being augmented.¹ Maintaining clinical expertise may prioritize humans performing clinical tasks, leveraging AI in the background to help confirm the physician's inputs prior or to point out opportunities, like re-reviewing images.
• Build AI literacy into training. Early-career physicians and residents are the most vulnerable to never-skilling. Education on AI’s limitations, including its pattern recognition without context and its confidence without lived clinical experience, can be woven into residency and CME, not treated as a technology orientation.
• Design for transparency. Patients may not see the AI operating in the background, but they notice when their physician is fully present. A 2025 athenahealth and United States of Care survey found that 58% of physicians worry about overreliance on AI for diagnosis. That concern is a healthy instinct — and health systems and technology vendors should design for it, not around it. 

The goal isn't less AI... it’s better-positioned AI 

The deskilling fear is real, documented, and clinically significant. The Lancet colonoscopy study isn’t a case against AI in medicine. Rather, it’s a case for intentional design and deployment, both immediately and for the long-term. The research points to the ongoing need for AI that genuinely augments a physician’s judgment, that keeps the human actively in the loop on decisions that matter, that supports skill development rather than substituting for it.

For EHR and health IT vendors, this is both a design responsibility and a differentiation opportunity. The platforms that earn long-term trust from clinicians will be the ones that make physicians better at their jobs. What “better” looks like in the future will be different than it was in the past as the way medicine is practiced continues to shift. But the goal must remain enhancing and skilling for this new AI-augmented reality.