Exactitud y coherencia de las respuestas de diferentes plataformas de inteligencia artificial sobre el manejo de accidente de hipoclorito de sodio.

Abstract

Problem: Large language models (LLMs) are emerging as complementary tools in endodontics, but their reliability in handling certain topics has not been thoroughly evaluated. Objective.- To assess the accuracy and consistency of artificial intelligence responses regarding the management of sodium hypochlorite accidents in endodontics. Methodology.- A quantitative, crosssectional study evaluated ChatGPT- 4, ChatGPT-4o mini, DeepSeek AI (V3), Google Gemini (Flash 2.5), and Perplexity AI by submitting 33 questions on sodium hypochlorite accidents, repeated three times per day over seven days. The questionnaire included 11 multiple-choice items, 11 fill-in-the-blank items, and 11 true/false items. Results.- A total of 3,465 responses were recorded and coded as correct or incorrect. Data were analyzed using intraclass correlation coefficients and chi-square tests (α = 0.05). All LLMs showed high repeatability (ICC 0.857–0.966; p < 0.001). Overall accuracy did not differ significantly across platforms (χ²(4) = 62.942; p < 0.001), but varied by question format (p < 0.05), with lower accuracy observed in true/false and fill- in-theblank items. Google Gemini achieved approximately 80% accuracy, while ChatGPT-4o mini and Perplexity AI exceeded 90% in multiple-choice questions. Gemini's explanations contained errors, and its references lacked evidentiary value. Conclusion: Although LLMs show potential as auxiliary tools, human oversight remains necessary.