An official website of the United States government
Unlocking the Ability of Generative Artificial Intelligence to Improve Health Communication and Health Education
ORIGINAL RESEARCH — Volume 23 — October 28, 2026
PEER REVIEWED
The lowest median RGL of any prompt, 6.6, came from prompt 25, above the desired fifth- to sixth-grade target for complexity level 1. However, by generating 36 outputs for each condition, target RGLs, including for level 1, were predominately achieved. The median (IQR) RGLs were 5.8 (5.4–6.4) for complexity level 1, 7.5 (7.4–7.7) for complexity level 2, 9.5 (9.4–9.7) for complexity level 3, and 11.5 (11.3–11.7) for complexity level 4, with significant differences between complexity levels.
Figure 1. Reading grade levels for each complexity level. Boxes indicate the 25th percentile, median, and 75th percentile; dots indicate outliers; whiskers indicate 1.5 IQR.
Word counts for selected outputs were 97.0 (75.0–166.3) for complexity level 1, 258.0 (177.0–337.3) for complexity level 2, 345.0 (286.8–395.5) for complexity level 3, and 357.0 (318.8–401.0) for complexity level 4, with significant differences between complexity levels.
Figure 2. Word counts of optimized outputs for each complexity level. Boxes indicate the 25th percentile, median, and 75th percentile; dots indicate outliers; whiskers indicate 1.5 IQR.
Accuracy edits for raw LLM outputs were minimal, with 93.5%, 95.5%, 95.5%, and 95.0% of outputs for complexity levels 1 through 4, respectively, requiring no word additions or deletions. For outputs that required edits, median (IQR) word additions were 6.0 (2.8–11.0) for level 1, 4.0 (1.5–5.5) for level 2, 5.0 (1.0–9.0) for level 3, and 3.0 (2.0–6.0) for level 4. Word deletions were 8.0 (5.0-13.5) for level 1, 10.0 (5.3–11.8) for level 2, 4.5 (1.8–22.3) for level 3, and 5.0 (12.5–20.8) for level 4.
Figure 3. Jitter plot of the number of modifications made during expert review of medical information generated by prompts submitted to ChatGPT. Dots indicate the number of words reviewers deleted from and added to selected outputs. The short, thick lines at zero indicate the outputs that required no edits.
The opinions expressed by authors contributing to this journal do not necessarily reflect the opinions of the U.S. Department of Health and Human Services, the Public Health Service, the Centers for Disease Control and Prevention, or the authors’ affiliated institutions.