Artificially intelligent language models achieved an average accuracy of 81.2 percent when generating procedural descriptions for peripheral nerve surgery, but performance varied considerably and declined sharply when models were given simple rather than structured prompts. In an investigation published in Global Surgical Education, researchers evaluated how effectively advanced text-generation systems could delineate specialized operative workflows for surgical trainees.
Self-directed learning has become increasingly critical within surgical education, prompting exploration into whether large language models can reliably assist trainees. However, the precision, consistency, and vulnerability of these systems to differing model generations and prompt formulations have remained uncertain.
The study established reference standards by randomly selecting fifteen peripheral nerve procedures from the Washington University in St. Louis Surgical Education Website and coding each operation into individual surgical steps. The team then assessed GPT-5.2 across 90 generated responses using a structured prompt to evaluate procedural accuracy alongside other response characteristics. Linear regression models tested whether accuracy correlated with procedural complexity, total word count, overall step count, or step density. Robustness was subsequently examined in the newer GPT-5.5 model across five surgical procedures tested with both structured and simple prompts.
Across all 90 evaluations, GPT-5.2 achieved a mean accuracy of 81.2 percent with a standard deviation of 11.5 percent, with individual scores spanning from 68.5 percent to 91.7 percent. Statistical analysis revealed that while accuracy was positively associated with output word count and negatively associated with step density, neither metric remained an independent predictor of accuracy in regression modeling.
A detailed examination of errors identified 195 total procedural steps missed across the evaluations. The vast majority of omitted details involved higher-order clinical considerations: 86 of the 195 missed steps entailed judgement-based maneuvers, 56 represented major operative steps, and 41 involved critical neurovascular identification steps.
When evaluated across five procedures using structured prompts, GPT-5.5 yielded a comparable mean accuracy of 81.4 percent with a standard deviation of 16.2 percent, demonstrating greater performance variance than its predecessor. When prompted with simple instructions, however, GPT-5.5 accuracy dropped precipitously to 40.5 percent with a standard deviation of 40.2 percent, shifting procedural omissions directly toward major surgical steps. The authors concluded that while these models may serve as helpful tools for preliminary surgical orientation, their marked sensitivity to prompting strategies and unreliable coverage of judgement-dependent steps preclude them from replacing expert guidance during safety-critical operative training.
Subject of Research: Accuracy and prompting sensitivity of large language models describing peripheral nerve surgery procedures
Article Title: Accuracy of AI-generated procedural descriptions in peripheral nerve surgery: implications for surgical training
Article References: Chai, B., Yim, M., Ranjbar, D., & Tang, D. T. (2026). Accuracy of AI-generated procedural descriptions in peripheral nerve surgery: implications for surgical training. Global Surgical Education – Journal of the Association for Surgical Education, 5(1), Article 217. https://doi.org/10.1007/s44186-026-00626-z
Image Credits: AI Generated
DOI: 10.1007/s44186-026-00626-z
Keywords: peripheral nerve surgery, surgical education, ChatGPT, GPT-5.2, GPT-5.5, large language models, procedural descriptions, surgical training, prompting strategy, structured prompts, simple prompts, judgement-based maneuvers, neurovascular identification, step density, Washington University in St. Louis, Global Surgical Education, procedural accuracy, self-directed learning
News Source: Ophelia Keating. (November 12, 2026). ChatGPT Accurately Outlines Peripheral Nerve Procedures but Falters Without Structured Prompts. Scienmag.



