Assessment of patient information quality provided by artificial intelligence-based large language models on cryptorchidism: a comparison of ChatGPT-5, Gemini, and Grok

SUMMARY

Objective

The aim of this study was to evaluate the responses generated by ChatGPT-5, Gemini, and Grok to the “six most frequently asked patient questions” on cryptorchidism published by the European Association of Urology, assessing them in terms of quality, understandability, actionability, and readability.

Methods

ChatGPT-5, Gemini, and Grok were asked these six frequently asked questions listed on the European Association of Urology patient information page on cryptorchidism. The quality of the responses was evaluated using the Quality Assessment Tool for Patient Health Information instrument, understandability and actionability were assessed using Patient Education Materials Assessment Tool scores, and readability was measured with the Coleman-Liau Index. All evaluations were performed by four urologists.

Results

Among the three artificial intelligence-based large language models, Gemini achieved the highest mean Quality Assessment Tool for Patient Health Information and Patient Education Materials Assessment Tool for Printable Materials scores. Kruskal-Wallis analysis demonstrated a statistically significant difference in Quality Assessment Tool for Patient Health Information scores among the groups (p=0.005); pairwise comparisons revealed that Gemini scored significantly higher than ChatGPT-5 (p=0.001). No significant differences were observed among the models for Patient Education Materials Assessment Tool—Understandability Section or Patient Education Materials Assessment Tool—Actionability Section scores (p>0.05). In the readability analysis, Grok had the highest Coleman-Liau Index value (Coleman-Liau Index=14.68), and all models produced texts requiring a university-level reading ability. Although the Gemini model achieved higher overall quality scores, all artificial intelligence-based large language models provided good-quality but difficult-to-read information.

Conclusion

The responses generated by all three models demonstrated high levels of understandability and strong actionability. We anticipate that future, more advanced versions of artificial intelligence-based large language models will further improve these outcomes and contribute positively to the existing literature.