The quality, accuracy, and readability of information about hidradenitis suppurativa provided by artificial intelligence: Comparative analysis of ChatGPT, Copilot, and Perplexity

Abstract

Owing to shame and stigmatization, hidradenitis suppurativa (HS) patients may seek information about their disease on artificial intelligence (AI) chatbots. We aimed to evaluate the readability, quality, and accuracy of HS-related information provided by 3 AI chatbots: ChatGPT-4o, Copilot, and Perplexity. The 24 most frequently queried keywords regarding HS were identified using Google Trends. In this observational and cross-sectional study, we asked ChatGPT-4o, Copilot, and Perplexity chatbots for these keywords. The readability was evaluated using Flesch readability ease and Flesch-Kincaid grade level scores. SpaCy software v3.8.2, TERA (The Text Ease and Readability Assessor), TAALED (Tool for the Automatic Analysis of Lexical Diversity, βv1.4.1), and TAALES (Tool for the Automatic Analysis of Lexical Sophistication, v2.2) were used for the further linguistic analysis. The ensuring quality information for patients (EQIP) and DISCERN tools were used to assess the quality. The accuracy was assessed using a 6-point Likert scale. Perplexity exhibited the highest text length (P < .001). Copilot exhibited better readability scores (P < .001). Perplexity had higher FKGL (P = .001) and lower FRES (P = .001) than the others regarding the outputs in the "test, operation, investigation, or procedure & drug, medication, or product" category. Nevertheless, none of the chatbots achieved the necessary level of readability. Lexical diversity was lower in Perplexity than in the other chatbots (P < .001). Referential cohesion was highest in Perplexity, whereas deep cohesion was highest in Copilot (P < .001 and P = .009, respectively). Age of acquisition was lowest in Copilot responses (P < .001). Copilot achieved the highest EQIP score with "good quality with minor problems" (P < .001). ChatGPT had the lowest DISCERN scores (P = .001). Although all chatbot models displayed favorable accuracy results, Perplexity had higher accuracy scores compared to Copilot (P = .020). All the AI models had difficult reading levels as college or postgraduate. Copilot seemed to generate lexically simpler outputs, while Perplexity produced longer responses with more referential cohesion but lower lexical diversity. ChatGPT, Copilot, and Perplexity seem insufficient for providing extensive, easily understandable, and exactly accurate medical information about HS.