Back

Evaluating Large Language Models for ADHD Education: A Comparative Study of ChatGPT 5, DeepSeek V3, and Grok 4

HAN, X.; Xing, R.; Zhou, M.

2025-10-14 sports medicine
10.1101/2025.10.12.25337841 medRxiv
Show abstract

BackgroundChildren with attention-deficit/hyperactivity disorder (ADHD) often face barriers to participating in organized sports, particularly when physical education (PE) is delivered by outsourced coaches with limited training in disability inclusion. Meanwhile, large language models (LLMs) such as ChatGPT, DeepSeek, and Grok are increasingly used to generate educational content, yet their readability, stability, and accuracy for non-specialist educators remain unclear. MethodsThis study systematically compared three advanced LLMs, ChatGPT-5, DeepSeek V3, and Grok 4, using identical prompts related to ADHD definitions, symptoms, and medication-exercise interactions. Thirty responses per model were collected and analyzed for content accuracy, readability (Flesch-Kincaid Reading Ease, Grade Level, and SMOG), and lexical complexity. ResultsAll models aligned with DSM-5 in describing ADHD but differed in emphasis and stability. DeepSeek V3 produced the broadest and most variable outputs, Grok 4 showed the greatest consistency and clinical structure, and ChatGPT-5 generated concise and strengths-based explanations. However, all models exhibited high reading levels (FKGL > 12), exceeding recommended public-health standards. ConclusionWhile LLMs demonstrate strong potential for generating ADHD-related educational materials, their current readability and stability limitations restrict accessibility for non-specialist educators. Future work should focus on optimizing prompt design and language calibration to enhance usability in inclusive education contexts.

Published in PLOS ONE (predicted rank #1) · training set

Matching journals

The top 3 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.