Plast Reconstr Surg Glob Open. 2026 Mar 24;14(3):e7539. doi: 10.1097/GOX.0000000000007539. eCollection 2026 Mar.
ABSTRACT
BACKGROUND: Nearly 90% of Americans face health literacy challenges, limiting their ability to understand complex medical information. In plastic and reconstructive surgery, much of the literature exceeds recommended readability levels, creating barriers to patient education, including disparities-focused research. This study evaluated whether large language models (LLMs) can generate accurate and patient-accessible summaries of such research.
METHODS: Eight disparities-related plastic surgery articles from PubMed were input into 4 LLMs: ChatGPT-4o, Gemini 1.5 Pro, Grok 3, and DeepSeek-V3, using a standardized prompt to simplify the text to a sixth- to eighth-grade reading level. Generated summaries were assessed using 4 readability metrics (Flesch Reading Ease [FRE], Flesch-Kincaid Grade Level [FKGL], Simple Measure of Gobbledygook Index, and Gunning Fog Index) and were reviewed by a physician for accuracy.
RESULTS: Grok generated the most readable summaries, achieving average scores between the seventh- and ninth-grade reading levels (FRE M = 63.06, SD =1.80; FKGL M = 7.69). It significantly outperformed the other models across all metrics (FRE P = 0.001; FKGL P = 0.003; Simple Measure of Gobbledygook P = 0.034; Gunning Fog Index P = 0.007). ChatGPT, Gemini, and DeepSeek showed moderate improvements but did not achieve statistically significant differences from the original articles (P > 0.05), with average grade levels between the 10th and 12th grades.
CONCLUSIONS: Grok demonstrated superior readability while preserving accuracy, making it the only LLM to meet health literacy benchmarks. Other models fell short, underscoring a gap in artificial intelligence tools. Enhancing LLM performance could promote access to surgical literature and empower diverse patient populations through enhanced health communication.
PMID:41884761 | PMC:PMC13012324 | DOI:10.1097/GOX.0000000000007539
