This summer, the American Gastroenterological Association (AGA) introduced Nigel, a point-of-care AI assistant built specifically for gastroenterology and hepatology. It takes its place beside gastroenterology-specific tools such as Medtronic Inc’s GI Genius, an AI module that enhances detection of colorectal polyps in real time during colonoscopy and has been shown to outperform standard colonoscopy.

General medicine platforms commonly used by gastroenterologists include, among others, DoxGPT, ChatGPT, Open Evidence, Claude, Gemini, and Microsoft Copilot, often alongside traditional resources such as UpToDate and PubMed. AI applications are poised to become the mandated standard of practice in gastroenterology, and in some institutions already are. And you might say that the current explosion of intelligent products aside, AI has been broadly available to assist physicians since computer-aided screening and diagnosis were FDA approved for mammography in 1998.

Nigel, unlike tools trained on broad medical internet data, has GI-specific advantages: it’s built upon the AGA’s peer-reviewed clinical guidelines, structured around the GRADE evidence framework, and vetted by a panel of GI and liver experts to ensure clinical accuracy. According to its developers, Dennis L. Shung, MD, MHS, PhD, Bradly C. Stadie, PhD, and Aly T. Strauss, MD, MIE, the next-gen virtual platform provides “transparent, guideline-informed clinical guidance that notes the strength of the available evidence and areas of uncertainty.”

Nigel functions as a speedy virtual GI colleague and second opinion. “Nigel is different because it was built specifically for GI and liver medicine, is grounded in AGA guidelines, and was developed and vetted by GI and liver experts,” Strauss, an assistant professor of medicine at Johns Hopkins University in Baltimore, told Medscape Medical News. “The goal isn’t just to give an answer…it’s to give an answer that’s grounded in the evidence and guidelines gastroenterologists already trust. General AI knows a little about everything. Nigel was designed to know GI exceptionally well,” Strauss added.

As to feedback so far, she said, “GI clinicians appreciate having a specialty-specific tool they can use at the point of care. Clinicians don’t necessarily want another AI chatbot. They want an AI colleague that speaks the language of GI.”

Whatever AI’s shining promise, gastroenterologists have many concerns about its actual benefits vs its potential harms.

The following informal sampling of opinion reflects the hopes and fears of three prominent gastroenterologists: Xavier Llor, MD, PhD, a professor and director of the GI and Pancreatic Cancer Prevention Program at Yale School of Medicine in New Haven, Connecticut; Benjamin Lebwohl, MD, MS, a clinical researcher at Columbia University Irving Medical Center in New York City; and Nabil Mansour, MD, an associate professor and director of the McNair General GI Clinic at Baylor College of Medicine in Houston.

Llor

To date, no definitive head-to-head efficacy comparisons have been made between different AI instruments in gastroenterology, so how should gastroenterologists choose a particular AI product from the options available?

“Gastroenterologists should understand who developed the product, what data it was trained and validated on, how it performs according to accepted clinical and scientific standards, and whether it’s been independently evaluated,” Llor told Medscape Medical News.

He noted a major challenge is that the rapid evolution of the field means that the traditional scientific model for assessing performance is difficult to apply in real time. “For that reason, independent validation and post-implementation monitoring will be essential. Over time, we will need comparative effectiveness data to understand which tools perform best, in which settings, and for which clinical tasks.”

In Llor’s view, AI tools have tremendous potential and will doubtless help improve patient care. “However, we need to develop mechanisms that allow us to continuously assess their accuracy, safety, limitations, and real-world impact,” he cautioned. “The key will be to use these tools thoughtfully, as a way to strengthen clinical care rather than simply add another layer of technology.” 

Lebwohl

“Our hospital system already has CADe (computer-aided detection) installed in all units so that gastroenterologists can now perform AI-assisted colonoscopy,” Lebwohl said. “This technology, in which AI image detection points out potential polyps in real time, has been shown to increase the detection rate of precancerous polyps. We have reason to believe that removal of more precancerous polyps makes colonoscopy more effective at preventing colorectal cancer.”

His hospital system has also made OpenEvidence, a population-based general medical AI platform, available to all clinicians in its electronic health record. “This tool has been widely adopted by clinicians for providing answers to clinical questions and citing the relevant literature.”

Considering the growing availability of virtual second opinions, Lebwohl cautioned that “deskilling” is a concern in procedure performance and interpretation of the medical literature when physicians rely on AI in colonoscopy, as well as in evidence review. “This is a rapid area of research, and our medical school and training programs are engaging closely with this challenge,” he said. “We want to strike a balance between imparting fundamental skills to physicians and providing them with the latest tools they can use to practice medicine at the highest level.”

Other major safeguards that must be in place are accuracy — and privacy. “Though the large-language models continue to improve, they are not infallible, and require critical judgment to be used properly,” he said. “Clinicians and trainees need to understand how to use these technologies in a way that is compliant with medical privacy laws; it is key for hospitals and other employers of clinicians to recognize this, and to provide tools that can be safely used in this regard.”

Mansour

“Many clinicians here at Baylor are beginning to use AI-enabled clinical reference tools such as OpenEvidence and DoxGPT in their day-to-day practice,” said Mansour. “Because adoption varies among individual physicians, I would distinguish this from a formal, department-wide implementation. These tools are generally used as a sounding board, similar to informally consulting a colleague for a second opinion.”

They can help clinicians rapidly review a differential diagnosis, identify relevant literature, clarify guideline recommendations, consider alternative management strategies, or organize complex clinical information. Their main advantage is speed: They can synthesize a large amount of medical information within seconds and may draw attention to possibilities or evidence that the clinician has not yet considered.

“However, they are not relied upon as the sole source for major or critical medical decisions. Their output may be incomplete, outdated, overly confident, or based on evidence that does not fully apply to the individual patient,” he said. “They can also possibly generate inaccurate information, omit important nuances, or fail to account for comorbidities, patient preferences, local practice patterns, and other contextual factors.” Even when an answer from an AI tool appears well referenced, the clinician must confirm that the cited evidence is accurate and appropriately interpreted, Mansour cautioned.

In his view there’s a risk these tools could be embraced too eagerly — “particularly if convenience is mistaken for reliability.” They should not replace a careful history, physical examination, review of the medical record, direct evaluation of the available data, or consultation with an appropriate specialist. “There is also a potential for automation bias, in which clinicians may give undue weight to an AI-generated recommendation simply because it is presented clearly and confidently,” he said.

According to Mansour, the most appropriate model is one in which AI augments rather than replaces human judgment. “The clinician remains responsible for evaluating the quality of the information, recognizing when the AI may be wrong, integrating the patient’s specific circumstances, and making the final decision,” he said. “Used in that manner, these tools can broaden a clinician’s thinking and improve efficiency without displacing critical judgment.”

Another area where more formal adoption is growing is in ambient AI documentation. “Baylor is currently piloting ambient AI scribes, which use speech recognition and large language models to generate draft clinical notes from the conversation between the physician and patient. Rather than replacing the clinician, these systems are designed to reduce the documentation burden by automatically producing structured notes that the physician reviews, edits as needed, and approves.”

Early experience suggests these can reduce time spent on documentation and allow clinicians to focus more of their attention on the patient during the visit. “As with any AI-generated content, however, the draft note must be carefully reviewed for accuracy and completeness before it becomes part of the medical record.”

CADe during colonoscopy using the GI Genius system is “a different type of point-of-care AI. Rather than providing diagnostic or treatment recommendations, the system analyzes the live colonoscopy video and highlights areas that may represent polyps,” said Mansour. “Its purpose is to serve as an additional set of eyes and potentially reduce missed lesions. The endoscopist must still determine whether the highlighted area is truly abnormal, characterize the lesion, decide whether it should be removed, and select the appropriate resection technique.”

Overall, AI is likely to become an increasingly routine part of clinical care. Its greatest near-term value may be in improving efficiency, expanding access to information, and providing an additional layer of support. “The central challenge is ensuring that adoption is accompanied by appropriate validation, transparency, clinician oversight, and recognition of the technology’s limitations,” Mansour said.

The consensus is that while the road ahead holds great promise of clinical benefits, there are twists and turns that must be negotiated by the irreplaceable judgment of skilled physicians.

Strauss is one of the developers of Nigel but had no other relevant conflicts of interest. Llor and Lebwohl had no relevant conflicts of interest with regard to their comments. Mansour reported being a consultant for Medtronic Inc.