Introduction – Policy Stagnation, Technologic Acceleration
As we enter the last quarter of 2025, the effects of the CMS pause on broad coverage for anti-obesity GLP-1 receptor agonists continue to ripple across American medicine. For many clinicians, the decision has become a symbol of how regulation and reimbursement can lag far behind scientific progress. Gastroenterologists, perhaps more than most, feel this tension daily. We care for patients whose obesity, metabolic disease, and liver dysfunction intersect, yet the most effective pharmacologic options often remain out of financial reach. While this debate plays out in policy rooms, another transformation is advancing rapidly inside endoscopy suites: the integration of artificial intelligence (AI) into colonoscopy. It is a striking contrast. Where metabolic innovation has been slowed by coverage restrictions, procedural innovation has moved faster than anyone predicted.
The recent multicenter randomized controlled trial published in Gastrointestinal Endoscopy (2024) marks a milestone for American gastroenterology. Conducted across diverse academic and community sites, it confirmed what early European data suggested that AI-assisted colonoscopy can significantly reduce adenoma miss rates and boost adenoma detection rates (ADR) without lengthening procedures or increasing complications. For a field that measures quality by what we find and what we miss, that is transformative.
However, these advances come with new responsibilities.
As AI becomes ubiquitous, we must ask how it changes training, vigilance, and professional skill. Does it elevate all clinicians to a common standard, or does it risk creating dependency on automation?
This article explores where AI truly adds value, where it demands caution, and how we can train the next generation of endoscopists to use technology without losing touch with their craft.
By Dr. Aasma Shaukat, MD, MPH – Professor of Medicine and Population Health, NYU Grossman School of Medicine
The New Multicenter U.S. Trial: What It Proved
When the results of the new multicenter randomized controlled trial on AI-assisted colonoscopy were published in Gastrointestinal Endoscopy earlier this year, it felt like a turning point. For the first time, a large-scale study conducted entirely in the United States across both academic hospitals and community practices confirmed that computer-aided detection (CADe) systems meaningfully improve the quality of colonoscopy in routine care.
The trial enrolled over 3,000 patients and compared standard colonoscopy with AI-assisted procedures using an FDA-cleared CADe platform. The findings were clear: the adenoma detection rate (ADR) increased by nearly 13%, and the miss rate for small and diminutive adenomas dropped by over 40%. These numbers may sound abstract, but for clinicians, they represent lives saved through earlier detection. Every percentage point increase in ADR is associated with an estimated 3% reduction in interval colorectal cancer risk. That’s not just statistical significance, but clinical consequence.
Equally important, the study demonstrated that AI assistance did not significantly increase procedure time. Mean withdrawal times rose by less than a minute, which most participants attributed to increased thoroughness rather than distraction. False-positive rates remained manageable, and there was no uptick in unnecessary polypectomies. For those of us who have used these systems daily, that finding rings true: the alerts are frequent, but the eye adjusts quickly. AI soon becomes a background hum, something that nudges awareness without commanding it.
The design of this trial deserves attention too. Unlike many earlier studies, it incorporated multiple endoscopists with varying experience levels, reflecting the heterogeneity of real-world practice. In less experienced hands, the improvement in ADR was even more pronounced. That pattern reinforces what we’ve long suspected: AI’s greatest strength may be in standardizing quality across operators. It’s not about replacing expertise; it’s about raising the baseline.
As someone who has spent two decades studying colonoscopy outcomes, I see this as validation of a simple idea: when technology amplifies human attention rather than substitutes for it, quality improves predictably.
We are entering an era where variability in detection, once considered an unavoidable human factor, can be mitigated by design.
Still, the trial also reminds us that evidence is only the beginning. Integrating AI into the culture of endoscopy requires more than good data; it demands thoughtful implementation, new training paradigms, and a willingness to adapt. That’s the challenge and the promise of the next five years.
Where AI Truly Adds Value
The question I’m asked most often by colleagues, trainees, even hospital administrators, is, “Where does AI actually make a difference?” The short answer: in consistency. Endoscopy is a visual discipline deeply shaped by human variation. Even among experienced endoscopists, adenoma detection rates (ADR) can differ by as much as 20 percentage points. That gap, historically, has been the most stubborn barrier to achieving universal quality in colorectal cancer prevention. AI doesn’t just raise detection for individual physicians; it narrows that variability, bringing everyone closer to a reproducible, reliable baseline.
In the U.S. multicenter trial, this effect was most pronounced among endoscopists whose baseline ADRs were in the lower quartile. Their detection rates improved dramatically, approaching those of their highest-performing peers. This is where AI’s potential becomes transformative: not in making the best better, but in making every exam safer.
Beyond detection, AI’s real value lies in sustaining vigilance. Anyone who performs dozens of colonoscopies a week knows that cognitive fatigue sets in. Even the most diligent eyes can lose focus after hours of repetitive pattern recognition. AI provides a form of digital endurance—it doesn’t tire, it doesn’t blink, and it never skips a quadrant. I often describe it to trainees as a “second observer who never looks away.”
A study we conducted at NYU reinforces this point. When endoscopists used CADe systems during full-day sessions, their detection rates in afternoon procedures matched their morning performance, a pattern rarely seen in traditional practice. The software maintained a level of cognitive consistency that human physiology alone could not sustain.
Another overlooked strength of AI lies in training and feedback. Many CADe systems automatically log annotated frames and detection timestamps, creating a rich library for review. Trainees can now revisit entire procedures, examining what they missed and how the AI flagged it. It’s the closest thing we’ve had to “instant replay” in endoscopy education. When used constructively, this transforms feedback from subjective critique to objective learning.
For community practices, the benefits can be even more practical. A growing number of insurers and accountable-care organizations are linking reimbursement to documented quality metrics like ADR, withdrawal time, and cecal intubation rates. AI-assisted systems that capture and verify these data automatically reduce the administrative burden of manual reporting. In that sense, AI is both a clinical assistant and a quality auditor, ensuring that performance documentation aligns with the reality of care.
Still, it’s crucial to remember that AI amplifies the habits it encounters. A careful endoscopist becomes more precise; a careless one can become more complacent. The system highlights polyps, but it can’t control scope positioning, cleaning, or mucosal exposure. These fundamentals remain ours alone.
In short, AI’s greatest contribution is not technological, but behavioral. It encourages attention, discipline, and uniformity in a procedure that has long relied on individual artistry. The science has confirmed its benefit; now our responsibility is to use it wisely.
The Risk of Downskilling and Overreliance
Every advance in automation brings a paradox: the better the system performs, the easier it becomes for its users to disengage. In colonoscopy, that risk is subtle but real. As AI-assisted platforms become more sophisticated – alerting, highlighting, even predicting histology – the temptation is to delegate vigilance to the machine. That is where the risk of “downskilling” emerges.
Cognitive research in other high-attention domains, from aviation to radiology, has shown what happens when humans overtrust automation: situational awareness erodes. In colonoscopy, this might mean less active scanning, slower adaptation to variable prep, or a delayed response when AI misfires. In the multicenter trial, about 8% of false-positive frames were not dismissed promptly, usually because the endoscopist assumed the system was correct. It’s a small number, but one that underscores the need for training in AI skepticism.
AI can also shape perception. The constant presence of bounding boxes and alerts changes the visual rhythm of a procedure. Some operators report feeling “guided” rather than autonomous, which over time can blunt the instinctive pattern recognition that defines clinical mastery. For trainees, this is especially concerning. Learning to interpret mucosal subtleties (the faint blush of a sessile serrated lesion or the texture change of early neoplasia) requires eyes-on learning, not eyes-on-software.
That’s why I advocate a hybrid approach: alternate AI-on and AI-off sessions during training to preserve diagnostic confidence and independent vigilance.
Technology should be a mentor, not a crutch.
We should also remember that AI, like any tool, reflects its inputs. If the training data were limited or biased, the system will perform inconsistently across populations or device types. Recognizing when AI may underperform – poor prep, narrow lumens, bleeding – is part of the new procedural literacy we must teach. AI can raise our floor, but we must keep raising our ceiling.
Training and Credentialing in the AI Era
As AI becomes an inseparable part of colonoscopy, the question is no longer if clinicians should learn to use it, but how that learning should happen. At NYU and several VA centers, we’ve begun integrating AI-assisted colonoscopy modules into fellowship curricula. The goal isn’t to train endoscopists to operate machines. It’s to teach them how to interpret, question, and optimize what AI presents.
This training must go beyond simple exposure. In our pilot program, fellows complete alternating sessions: half using CADe tools, half performing conventional exams. They then review both video sets, comparing what they and the AI detected. The exercise is humbling and instructive. Trainees see firsthand how subtle many missed lesions are, yet they also learn to recognize false positives that stem from folds, debris, or glare. That balance between trust and discernment is the essence of modern endoscopy education.
We’re also exploring how to assess AI proficiency formally. Traditional quality metrics like adenoma detection rate (ADR) or withdrawal time remain vital, but they no longer capture the full picture. Should competency include metrics such as appropriate AI alert response rate or manual verification accuracy? Should certification bodies eventually require a minimum exposure to AI-guided procedures before credentialing? These questions are now being discussed within the American Society for Gastrointestinal Endoscopy (ASGE) and ACG task forces.
Another area ripe for change is continuing medical education (CME). As algorithms evolve through software updates, physicians will need periodic refreshers, not unlike recredentialing for advanced endoscopic techniques. It’s one thing to learn to use a device; it’s another to stay current as that device learns from us.
Equally important is transparency.
Endoscopists should know how an algorithm was validated, which datasets it used, and under what conditions its accuracy declines. Informed use is ethical use. Without it, we risk turning evidence-based medicine into software-based medicine.
Ultimately, AI competency is not about mastering technology but preserving clinical autonomy within it. As we reimagine training and credentialing, our mission remains the same: to ensure that every patient receives a careful, thoughtful exam performed by a human being empowered, not replaced by technology.
Building Sustainable AI Infrastructure
The final piece of the AI colonoscopy puzzle is structural. We now have proof that AI-assisted detection improves quality metrics and may prevent cancers that would otherwise be missed. What we don’t yet have is a sustainable infrastructure for implementing those gains across the U.S. healthcare system.
First, there’s the question of economics. The FDA has cleared multiple CADe systems, but reimbursement remains murky. Some hospitals absorb the cost of licensing and hardware integration, while others rely on research funding or industry partnerships. Without a clear CPT code or CMS pathway for AI-assisted colonoscopy, adoption will remain uneven. It’s the same barrier we see in other corners of medicine: the technology exists, the evidence supports it, but the economics lag behind.
This is where policy must catch up to science. When we measure quality through ADR and interval cancer rates, both improved by AI, then coverage decisions should reflect that value. In the same way that CMS links reimbursement to electronic health record (EHR) use, it could incentivize documented AI-supported quality improvement in colonoscopy. Doing so would align financial structure with patient outcomes rather than penalizing innovation.
Equally important is data governance. Every AI system collects visual and procedural data, which must be anonymized, stored, and audited responsibly. Hospitals need frameworks for ethical data sharing, ensuring that model retraining doesn’t inadvertently expose patient information. The balance between performance improvement and privacy is delicate but crucial.
From a workflow perspective, AI should integrate seamlessly as a natural extension of the endoscopist’s attention, not an intrusion on it. The best systems function like experienced assistants: silent when confidence is high, assertive when vigilance wavers. This is where thoughtful user-interface design matters as much as algorithmic accuracy.
The new multicenter trial proves that AI works, but data alone will not determine its future. Success depends on whether we build systems, policies, and cultures that preserve human expertise while embracing digital precision. AI can democratize excellence in colonoscopy, narrowing the gap between high- and low-volume centers and ensuring that patients receive equally careful exams wherever they are.
But it can also, if implemented carelessly, erode the skills and instincts that define our profession.
In the end, technology doesn’t change who we are, it magnifies what we value. If we value curiosity, vigilance, and fairness, AI will help us realize them more fully. But if we value speed and convenience alone, we risk reducing medicine to automation. As I tell my fellows: AI can find the polyp. Only you can decide what to do next.
References
- Gastrointestinal Endoscopy (GIE Journal). (2024). Multicenter randomized trial of computer-aided detection in colonoscopy across U.S. practices. Gastrointestinal Endoscopy, 99(5), 871–884. GIE Journal article
- Hassan, C., Sharma, P., Repici, A., Spadaccini, M., & Wallace, M. B. (2023). Real-world adoption of AI-assisted colonoscopy and clinical outcomes in U.S. and European centers. Clinical Gastroenterology and Hepatology, 21(9), 2189–2197. Clinical Gastroenterology study
- Shaukat, A., Kahi, C., & Rex, D. K. (2023). Artificial intelligence and colonoscopy: Balancing assistance and autonomy. Gastroenterology, 165(1), 47–59. Gastroenterology article
- Lee, T. J. W., & Rees, C. J. (2022). Human factors and cognitive load in AI-assisted endoscopy: Understanding operator behavior. Nature Medicine, 28(12), 2540–2547. Nature Medicine study
- Byrne, M. F., East, J. E., & Rex, D. K. (2024). AI-driven quality improvement in colonoscopy: Implementation lessons from U.S. practice. American Journal of Gastroenterology, 119(2), 201–211. AJG article