How does ChatGPT respond to stuttering-related frequently asked questions? A mixed-methods, cross-version comparison

This study compared responses to stuttering-related frequently asked questions (FAQs) generated by two versions of ChatGPT (the freely accessible ChatGPT-4 and the subscription-based ChatGPT-4.5), evaluating content, emotional tone, structural organization, and readability.

A set of 34 FAQs was developed through a two-stage process. First, the authors created an initial pool of questions. Then, 13 adults who stutter rated and refined these items for relevance. Each FAQ was entered once into GPT-4 and GPT-4.5. Responses were grouped into five broad content domains for computational analysis; these domains were identified and refined through thematic analysis. Responses were analyzed using (a) reflexive thematic analysis, (b) computational emotional analysis, and (c) quantitative readability analysis.

Both versions produced highly overlapping thematic content, yielding six shared themes. However, systematic cross-version differences emerged in how this content was delivered. GPT-4.5 produced substantially shorter responses (≈63% reduction in length) that were more structured and directive, whereas GPT-4 provided longer and more narrative explanations. Despite increased concision, GPT-4.5 responses were modestly less readable. Emotional profiles were largely congruent across models, dominated by anticipation and trust in therapy and parenting domains, with more neutral affect for causes and variability.

ChatGPT functioned as a first-line explainer for stuttering-related FAQs. It provided generally accurate basic information and often used relatively non-stigmatizing language, while consistently emphasizing the importance of treatment, particularly early intervention for stuttering. Model iteration primarily reshaped how information was framed and accessed rather than substantially changing what was conveyed. These findings underscored the importance of clinician oversight, accessibility considerations, and ethical caution when considering the use of AI-generated content in public health communication about stuttering.