TOPLINE:

Large language models (LLMs) can assist clinicians in gastroenterology and hepatology by drafting notes, summarizing complex data, and managing patient messages. But because they can produce confident yet incorrect or biased outputs and may expose sensitive health information, they must be deployed proportionally to clinical risk, with mandatory human oversight, compliance with privacy laws, use of vetted medical sources, transparent disclosure, and continuous monitoring and governance.

METHODOLOGY:

• Gastroenterology and hepatology practices generate large volumes of unstructured text — clinic notes, procedure and pathology reports, imaging data, and discharge summaries — that must be synthesized against evolving evidence and patient‑specific factors, making the field well suited for LLM-assisted documentation and decision support; however, in high‑stakes situations, hallucinated or biased outputs can lead to patient harm.
• A researcher conducted a targeted review of peer‑reviewed studies and policy documents, prioritizing material on the use of LLMs, medical ethics, data privacy, cybersecurity, and regulation in the US and other countries; PubMed/MEDLINE was searched for records in English from January 2018 through January 2026 using terms related to LLMs, gastroenterology and hepatology topics, and governance keywords.
• Each intended use of LLMs was matched to a risk tier based on four dimensions: clinical proximity and actionability, severity and reversibility of potential harm, degree of autonomy and opportunity for human oversight, and sensitivity of patient information.
• The tiers ranged from tier 0 (no clinical data and nonclinical support) to tier 4 (autonomous clinical action); tools spanning multiple functions were classified at the level of their highest‑risk or most directly patient‑impacting function.

TAKEAWAY:

• LLMs showed utility for tier 0 education tasks and delivered high-value tier 2 outputs (clinician-reviewed documentation and referral summaries) that shortened triage time but required validation for omissions, bias, and performance drift; tier 3 decision support gave mixed results (sometimes matching guidelines, sometimes discordant), so such tools need tight scoping, links to supporting evidence, formal validation, and clear clinician accountability.
• Because AI language tools can make confident mistakes, miss important warnings, or repeat biased assumptions that do not fit individual patients, their clinical use requires human oversight, targeted testing, transparent disclosure with opt‑out where feasible, and clear institutional responsibility and clinician training.
• When AI tools process patient information, organizations must apply the minimum necessary principle; use appropriate vendor contracts; comply with data‑protection laws; recognize that de‑identification does not eliminate the re‑identification risk; treat decision‑support tools as potentially regulated medical device/health information technology products requiring validation, lifecycle controls, and transparency; and ensure clinicians retain responsibility so AI is not the sole basis for denials or major treatment decisions without a documented individualized review.
• To use LLMs safely, organizations need clear policies and oversight, basic technical protections such as trusted data sources and access limits, and strong privacy and legal safeguards.

IN PRACTICE:

“LLMs can help teams manage information overload and reduce clerical burden, but safe use depends on aligning capabilities with clinical risk, protecting patient data, validating performance across diverse populations, and sustaining governance throughout the deployment lifecycle,” the author wrote.

SOURCE:

The review was authored by Sahil Khanna, MBBS, Mayo Clinic, Rochester, Minnesota. It was published online in Therapeutic Advances in Gastroenterology.

LIMITATIONS:

This review did not discuss any limitations.

DISCLOSURES:

This review received no financial support. The author declared having no competing interests.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.