Author: Matthew Clanfield, et al.
Journal of Clinical Anesthesia, 2026
Large language models are increasingly being used by physicians and trainees to obtain rapid clinical and procedural information. This study evaluated whether three commonly used artificial intelligence systems could provide safe, accurate, and useful troubleshooting advice to anesthesia trainees performing regional nerve blocks.
The investigators compared ChatGPT, Microsoft Bing Chat/Copilot, and Google Gemini across seven regional anesthesia techniques. Ten anesthesiologists independently evaluated the responses for safety, accuracy, and appropriateness for trainees.
Methods
The three large language models were asked to provide technical advice when a planned regional anesthetic block was unsuccessful.
Responses involving seven different nerve blocks were assessed. The evaluators examined:
• Accuracy of the procedural advice
• Safety of the recommendations
• Appropriateness for anesthesia trainees
• Readability and complexity
• Accuracy and existence of cited references
Interrater agreement was high, with a Cronbach alpha of 0.98.
Key findings
ChatGPT achieved an average score of 3.71 out of 5 across the seven regional anesthesia techniques.
ChatGPT scored higher than Gemini for every nerve block evaluated.
ChatGPT also performed better than Bing for troubleshooting femoral nerve blocks.
The overall quality of ChatGPT’s advice was considered reasonably useful, but it did not achieve a level that would support independent or unsupervised clinical use.
Readability
ChatGPT’s responses were written at approximately a university-graduate reading level.
This level of complexity may be appropriate for anesthesia trainees but may reduce accessibility for less experienced learners or other members of the healthcare team.
Reference accuracy
The most important weakness involved the references supplied by the models.
All references provided by Bing and Gemini were found to be genuine.
Only 25% of the references provided by ChatGPT actually existed.
This finding demonstrates that an artificial intelligence response can appear authoritative and provide clinically reasonable information while supporting it with fabricated or unverifiable citations.
Clinical implications
Large language models may provide useful suggestions when a regional block is incomplete or unsuccessful. Potential applications include:
• Reviewing relevant anatomy
• Suggesting adjustments in needle position
• Recommending alternative approaches
• Identifying causes of incomplete sensory distribution
• Reviewing local anesthetic dosing considerations
• Helping trainees organize a troubleshooting strategy
However, the information should be used only as an adjunct to clinical judgment, ultrasound findings, patient assessment, and direct supervision from an experienced anesthesiologist.
An LLM cannot evaluate the patient, view the ultrasound image, feel resistance during injection, verify needle-tip location, or recognize subtle signs of local anesthetic systemic toxicity, nerve injury, or intravascular injection.
Patient-specific decisions should therefore not be based solely on AI-generated recommendations.
Educational implications
Large language models could function as an immediately available supplementary teaching resource, particularly when a faculty member is not standing directly beside the trainee.
They may help trainees formulate questions, review alternative techniques, and think systematically about why a block failed.
The technology should not replace supervised regional anesthesia education, formal anatomical training, simulation, or expert bedside instruction.
Trainees should also be taught to verify all references and recommendations against trusted sources, including textbooks, society guidelines, peer-reviewed publications, and institutional protocols.
Important limitations
This was a single-center comparative study.
Only three large language models and seven nerve blocks were evaluated.
The quality of an LLM response depends heavily on the wording of the prompt and the specific version of the model used.
Large language models change frequently, meaning the results may not represent the performance of newer versions.
Anesthesiologists evaluated written advice rather than observing whether trainees could safely apply the recommendations to actual patients.
The study therefore did not establish that using an LLM improves block success, reduces complications, or improves patient outcomes.
Bottom line
ChatGPT provided the highest overall quality of troubleshooting advice among the three large language models evaluated, with a mean score of 3.71 out of 5.
Its advice was generally useful but was not consistently reliable enough for independent clinical decision-making.
The most concerning finding was that only 25% of the references supplied by ChatGPT were genuine.
Large language models may have a role as supplementary educational tools for regional anesthesia trainees, but their recommendations and citations must be critically evaluated and verified before clinical use.
Thank you to the Journal of Clinical Anesthesia for allowing us to summarize this article.