Glossary

Clinical LLM

What a clinical LLM is, how general models are adapted for healthcare work, what benchmark scores do and do not prove, and where regulators have drawn the lines. As of July 2026.

By Jonas WeirReviewed by Jonas Weir · editorial reviewUpdated

The short version

  • A clinical LLM is a large language model developed or adapted for healthcare tasks — clinical question answering, documentation, summarization — via domain training, fine-tuning, or grounding in clinical sources.
  • Benchmark progress has been steep: 67.6% on MedQA in 2023 (Med-PaLM era) to 86.5% with Med-PaLM 2.
  • Exam scores have yet to translate directly into practice: in a randomized trial, GPT-4 access moved physician diagnostic reasoning from 74% to 76% — a difference that did not reach significance.
  • WHO issued dedicated guidance on large multi-modal models for health in January 2024; FDA published draft lifecycle guidance for AI-enabled device functions in January 2025.

A clinical LLM is a large language model developed or adapted for healthcare tasks — answering clinical questions, drafting documentation, summarizing records — through domain-specific training, fine-tuning, or grounding in clinical knowledge sources, and evaluated against clinical benchmarks and clinician review 1.

Why do clinical LLMs matter in healthcare?

Language runs through almost every clinical workflow: notes, referral letters, discharge summaries, guideline lookups, patient messages. A model that handles clinical language well touches all of them at once, which is why clinical LLMs became the fastest-moving category of healthcare AI. The 2023 Nature paper that introduced Med-PaLM marked the turn: Flan-PaLM reached 67.6% on MedQA, the USMLE-style benchmark, beating the prior state of the art by more than 17 percentage points — while the same paper's human evaluation revealed gaps in comprehension, recall, and reasoning that the multiple-choice score concealed 1. That paper also set the evaluation template the field still uses: a composite benchmark (MultiMedQA, spanning six existing question sets plus consumer health searches) paired with clinician rating of long-form answers for factuality, potential harm, and bias — an early acknowledgment that no single number captures clinical fitness 1.

How does a clinical LLM work in practice?

Most clinical LLMs start life as a general foundation model and are specialized in one or more of three ways: instruction tuning on curated clinical question–answer sets, fine-tuning on domain text, or retrieval grounding in trusted sources at answer time. Med-PaLM 2 combined base-model improvements with domain fine-tuning and new reasoning strategies, reaching 86.5% on MedQA; in pairwise ranking, physicians preferred its long-form answers to physician-written answers on eight of nine clinical axes 2.

The scoreboard, though, is the easy part. The harder question is what happens when clinicians use these models. In a randomized clinical trial, physicians with GPT-4 access scored a median 76% per case on diagnostic reasoning versus 74% for those using conventional resources — an adjusted difference of 2 percentage points that did not reach statistical significance 3. Model capability and clinician benefit are separate claims, and each needs its own evidence.

Where do clinical LLMs appear today?

As of July 2026, clinical LLMs sit inside ambient documentation tools, patient message drafting, chart summarization, and clinical question answering — with benchmark results tracked in our LLM benchmark results tracker. Regulators have moved from commentary to frameworks. The WHO issued dedicated guidance on large multi-modal models for health in January 2024, carrying over 40 recommendations for governments, technology companies, and healthcare providers 4. In January 2025 the FDA published draft guidance on AI-enabled device software functions — lifecycle management plus marketing submission content — the frame under which LLM-based tools that meet the device definition would be reviewed 5.

Common misunderstandings

Exam performance equals clinical competence. MedQA measures multiple-choice recall under clean conditions. The Med-PaLM authors themselves paired their benchmark result with human evaluation precisely because the score alone overstated readiness 1.

A better model automatically helps the clinician. The randomized-trial evidence shows assistance effects can be small even when the model alone performs well — integration, prompting, and trust all mediate the outcome 3.

Clinical LLMs are unregulated. Use determines the pathway. Device-like claims pull an LLM tool into device review, and the FDA's 2025 draft guidance describes what submissions should contain across the product lifecycle 5. Confident fabricated text remains the failure mode to manage — see AI hallucination in clinical contexts.

Related terms

Questions & answers

  • Is a clinical LLM a separate kind of model from ChatGPT-style systems?

    Usually it is the same underlying architecture — a general foundation model — adapted for healthcare through instruction tuning on clinical question sets, fine-tuning on domain text, or retrieval grounding in clinical sources. Med-PaLM, the first widely studied example, was built by adapting a general model rather than training a new one from scratch.

  • Do high exam scores mean a clinical LLM is safe to use?

    No. Benchmark accuracy measures multiple-choice performance under clean conditions. In a 2024 randomized trial, giving physicians GPT-4 did not significantly improve their diagnostic reasoning versus conventional resources, and human evaluation of model answers still surfaces errors and omissions. Evaluation in the deployed workflow is the standard that matters.

  • Are clinical LLMs regulated?

    It depends on the claimed use. Tools that meet a medical-device definition fall under frameworks like the FDA's January 2025 draft guidance for AI-enabled device software functions; WHO's 2024 guidance on large multi-modal models addresses the wider governance questions for health uses that sit outside device regulation.

Sources

  1. Singhal K, Azizi S, Tu T, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172–180. doi.org/10.1038/s41586-023-06291-2
  2. Singhal K, Tu T, Gottweis J, et al. Toward expert-level medical question answering with large language models. Nature Medicine. 2025;31(3):943–950. doi.org/10.1038/s41591-024-03423-7
  3. Goh E, Gallo R, Hom J, et al. Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial. JAMA Network Open. 2024;7(10):e2440969. doi.org/10.1001/jamanetworkopen.2024.40969
  4. World Health Organization. Ethics and governance of artificial intelligence for health: guidance on large multi-modal models. Geneva: WHO; 2024. www.who.int/publications/b/70584
  5. U.S. Food and Drug Administration. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations; Draft Guidance. Federal Register. January 7, 2025. www.federalregister.gov/documents/2025/01/07/2024-31543/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing