Simpler diabetes advice was easier to read—but no AI prompt preserved every strength
In 384 model answers, a sixth-grade prompt improved readability while doctor-style prompts produced more comprehensive responses. The study tested text, not whether patients understood it or made safer decisions.
The short version
- What happened
- Researchers generated 384 answers to 16 type 2 diabetes-remission questions, crossing four language models with six prompting strategies; two experienced endocrinologists rated every response.
- Who it affects
- People seeking diabetes information need language they can understand without losing qualifications about eligibility, duration and medical supervision. This experiment identifies a design tension but does not show which wording helps a real patient make a better decision.
- What remains unproven
- Sixteen curated questions on one clinical concept
Keep in mind
Sixteen curated questions on one clinical concept
Synthetic responses only
Two clinician raters
No patient comprehension or outcome test
Readability formulas do not measure understanding
Exploratory correlations did not account for question clustering
Model versions will change
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Clearer language came with a measurable trade-off
A peer-reviewed study published on 29 October found no single way to prompt a language model that produced the strongest diabetes-remission advice on every measure. Researchers in China asked four systems—DeepSeek-V3.2, GPT-5.5, Gemini 3.1 Pro and Claude Sonnet 4.6—the same 16 core questions using six prompt styles. That balanced design produced 384 responses, all of which were included in the analysis.
The clearest contrast came from audience framing. The sixth-grade prompt reached a mean Flesch Reading Ease Score of 69.70, the best of the six strategies, while a doctor prompt scored 27.50, the hardest to read. Yet the sixth-grade prompt had the lowest descriptive composite quality score, 4.066 out of five, and the doctor prompt the highest, 4.350. Doctor prompts also had the highest mean comprehensiveness score, 4.609; patient prompts had the highest mean empathy score, 4.359.[1]
The comparison is useful, but it is not a patient trial
Two endocrinologists with more than five years of clinical experience independently scored accuracy, relevance, comprehensiveness, empathy and falsehood avoidance on five-point scales. The authors used mixed-effects models with the question as a random intercept. Model choice affected all five domains, although the reported effect sizes were mostly small or modest. GPT-5.5 had the highest numerical mean for accuracy and falsehood avoidance; DeepSeek led numerically on relevance, comprehensiveness and empathy. Those rankings are descriptive, not proof that one model is safest.
The study did not recruit patients, test understanding, measure treatment choices or observe clinical outcomes. Readability formulas count textual features; they do not establish that someone grasped the difference between remission and cure. The exploratory correlations also analysed responses without accounting for clustering by question, so they are hypothesis-generating. For clinicians and health-information teams, the practical lesson is to test both clarity and medical completeness with real intended readers, rather than treating a single prompt as a safety control.[1]
The distinction that matters
One prompt changed several qualities at once
The same prompt could make an answer easier to read while reducing other qualities in the study's rubric.
Sixth-grade prompt
Mean reading ease 69.70
Easiest to read; lowest composite quality
Doctor prompt
Mean reading ease 27.50
Hardest to read; most comprehensive
Patient prompt
Mean empathy 4.359/5
Highest empathy; not best on every measure
Funding and the next evidence
The work was funded by China's National Science and Technology Major Project, the China Academy of Chinese Medical Sciences and a hospital talent-development programme. The authors declared no competing interests. The publisher labels the paper peer-reviewed and citable, while noting that this accepted early version will receive further editing.
Confidence would rise with blinded evaluations involving more clinical topics, languages and independent patient groups; preregistered comparisons; and direct measures of comprehension, decision quality and harmful misunderstanding. Model versions also change, so this 2026 snapshot should not be treated as a permanent product ranking.[1]