How Do GenAI Chatbots Communicate Fluoride Information? A Readability and Negative Framing Analysis

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Introduction: As generative artificial intelligence (GenAI) chatbots increasingly mediate online health information seeking, this study aimed to assess the readability of chatbot-generated responses and characterize negatively framed content by estimating its prevalence and identifying its thematic patterns across different communicative scenarios and prompt formats. Methods: This cross-sectional infodemiological study evaluated fluoride-related responses generated by 5 freely accessible GenAI chatbots. To support this assessment, prompts were developed to simulate 3 communicative scenarios through which users may request or evaluate fluoride-related information: fact-checking, opinion, and explanation. Each prompt was formulated in both organic and structured versions, representing simple user-like questions and formulations expanded using prompt-engineering principles. Responses were segmented into sentences, which were independently classified by 2 trained and calibrated evaluators as positively/neutrally or negatively framed with respect to fluoride use for dental caries prevention. Readability was assessed using distinct indices, and topic modeling was performed to characterize thematic patterns. Results: A total of 1,326 sentences were analyzed, of which 1,040 were classified as positively/neutrally framed and 286 as negatively framed. The prevalence of negatively framed sentences did not differ significantly across chatbots or prompt formats. However, it varied according to the communicative scenario, with fact-checking prompts generating a higher prevalence of negatively framed content than opinion and explanation prompts. Topic modeling showed that negatively framed sentences were primarily organized around risk-oriented themes, including dental fluorosis, excessive fluoride exposure in children, toxicity, potential neurodevelopmental concerns, and skeletal fluorosis. Although structured prompts improved readability across most grade-level indices, chatbot-generated responses generally remained within difficult readability levels. Conclusion: GenAI chatbots have the potential to support communication about fluoride use; however, the frequent emphasis on fluoride-related risks in fact-checking responses and the generally difficult readability of the generated texts may hinder lay audiences’ interpretation of evidence-based recommendations. Knowledge Transfer Statement: Although generative artificial intelligence (GenAI) chatbots can provide adequate fluoride-related information, the presence of unfavorable content and high readability demands limit their practical usefulness. Because patients use these tools for health queries, clinicians must proactively address AI-driven misconceptions during consultations. Additionally, public health policymakers should advocate for digital oversight to ensure AI platforms deliver accurate, unbiased, and broadly accessible information regarding dental caries prevention.