Hybridization events in Squamata support the viviparity-driven conflict hypothesis.

BACKGROUND

The rise of large language models (LLM) promises to widely impact healthcare providers and patients alike. As these tools reflect the biases of currently available data on the internet, there is a risk that increasing LLM use will proliferate these biases and affect information quality. This study aims to characterize the effects of different race, ethnicity, and gender modifiers in question prompts presented to three large language models (LLM) on the length and readability of patient education materials about myopia.

METHODS

ChatGPT, Gemini, and Copilot were provided a standardized prompt incorporating demographic modifiers to inquire about myopia. The races and ethnicities evaluated were Asian, Black, Hispanic, Native American, and White. Gender was limited to male or female. The prompt was inserted five times into new chat windows. Responses were analyzed for readability by word count, Simple Measure of Gobbledygook (SMOG) index, Flesch-Kincaid Grade Level, and Flesch Reading Ease score. Significant differences were analyzed using two-way ANOVA on SPSS.

RESULTS

A total of 150 responses were analyzed. There were no differences in SMOG index, Flesch-Kincaid Grade Level, or Flesch Reading Ease scores between responses generated with prompts containing different gender, race, or ethnicity modifiers using ChatGPT or Copilot. Gemini-generated responses differed significantly in their SMOG Index, Flesch-Kincaid Grade Level, and Flesch Reading Ease based on the race mentioned in the prompt (p<0.05).

CONCLUSION

Patient demographic information impacts the reading level of educational material generated by Gemini but not by ChatGPT or Copilot. As patients use LLMs to understand ophthalmologic diagnoses like myopia, clinicians and users should be aware of demographic influences on readability. Patient gender, race, and ethnicity may be overlooked variables affecting the readability of LLM-generated education materials, which can impact patient care. Future research could focus on the accuracy of generated information to identify potential risks of misinformation.