AI Chatbots Deliver Surprising HIIT Information: All Four Miss Readability Benchmarks

A comparison of four AI services found clear differences in the quality and transparency of their exercise information, with every platform producing language above the researchers’ chosen reading-level target.

Asking a chatbot about high-intensity interval training can produce a confident, neatly organised explanation that demands more reading effort than a beginner might expect. A comparison of four AI services found clear differences in the quality and transparency of their exercise information, with every platform producing language above the researchers' chosen reading-level target.

The paper, "AI-generated high-intensity interval training information: a cross-sectional comparative evaluation of quality and readability," published in Frontiers in Public Health, examined responses from ChatGPT, Microsoft Copilot, Gemini and Perplexity. Researchers Zheng Zhang, Xiang-Yan Ruan and Kai-Tuo Zhu assessed how the answers performed on established information-quality measures and six readability formulas.

Perplexity received the highest median scores across all four quality instruments, and Copilot generally produced the simplest text. Those results describe specific features of the sampled answers. The researchers did not independently check factual accuracy, citation accuracy or clinical safety, making the findings a limited comparison rather than an endorsement of any platform's exercise guidance.

Four chatbots answered the same set of exercise queries

The researchers used worldwide Google Trends data covering August 20, 2021, to August 20, 2026, to identify topics connected with high-intensity interval training, commonly called HIIT. Screening reduced 29 candidate terms to 14 queries covering definitions, benefits, workouts, exercise examples, beginner information and Tabata. Broader concepts such as aerobic and anaerobic exercise were retained because of their relevance to HIIT explanations, without treating them as equivalent forms of training.

Each term was entered once into each chatbot in a fresh English-language conversation, producing 56 responses during collection beginning on August 20, 2026. The team preserved the original wording and used no follow-up questions, personalised instructions or requests for simpler language. Google Trends supplied a basis in relative search interest, rather than absolute search counts or evidence of how people actually phrase chatbot questions.

Two reviewers with doctorates in sports science independently assessed every response.

• DISCERN evaluated written consumer-health information;
• EQIP examined content, presentation and usability;
• The JAMA benchmark assessed authorship, attribution, disclosure and currency;
• GQS captured overall quality and usefulness.

The analysis accounted for the same queries appearing across platforms and adjusted follow-up comparisons to reduce the chance of overstating differences.

Higher quality scores came with important limits

Perplexity led on all four median scores, with its clearest relative advantage appearing in the transparency assessment. Copilot generally followed it on the other quality measures, and ChatGPT and Gemini received similar overall usefulness ratings. The table summarises the median scores, which represent the middle values across each platform's 14 responses.

Chatbot	DISCERN	EQIP (%)	JAMA transparency, out of 4	GQS, out of 5
ChatGPT	35.0	47.5	0	3.0
Copilot	39.0	57.5	0	3.5
Gemini	36.0	55.0	0	3.0
Perplexity	51.5	75.0	1	4.0

Under the paper's categories, Perplexity's DISCERN result sat near the bottom of the "good" range, and its EQIP result fell just below "excellent." Its transparency score reached only one of four available points. Perplexity scored higher than each rival on DISCERN, EQIP and JAMA in the adjusted comparisons; its GQS difference from Copilot did not clear the researchers' testing threshold.

A response can receive credit for presenting sources without those sources being accurate or supporting its claims. These instruments did not verify references, identify fabricated statements or establish whether a suggested workout would be medically appropriate, leaving those questions unanswered by the rankings.

Every platform missed the chosen reading-level target

The researchers assessed readability using six formulas and a prespecified sixth-grade benchmark for public-facing health information. All four platforms' median results missed the target across the evaluated indices. On the Flesch–Kincaid Grade Level measure, Copilot had the lowest median at 9.82, followed by Perplexity at 11.12, Gemini at 12.27 and ChatGPT at 14.40, placing even the simplest median response several school grades above the target.

The Flesch Reading Ease measure told a similar story: the study's target was at least 80, compared with medians of 43.94 for Copilot, 41.02 for Perplexity, 30.62 for Gemini and 6.47 for ChatGPT. Higher scores indicate easier text on this measure. The remaining formulas reinforced the reading-level gap, showing that higher quality scores did not consistently accompany simpler language.

Readability formulas estimate textual complexity, rather than directly testing what someone understands or can put into practice. Technical exercise terms and dense explanations could contribute to difficult scores, though the researchers did not test those possible explanations. The sixth-grade benchmark was a chosen evaluation target, rather than a universal boundary separating understandable information from inaccessible material.

A small snapshot leaves accuracy and practical use unresolved

The comparison covered four chatbots, 14 queries and one answer per query on each platform. Responses can change between attempts, and model updates, interfaces, retrieval functions and settings can alter the output. The results describe an August 2026 snapshot, with no assessment of repeatability, other languages or the longer conversations through which users might request clarification.

Important topics were not comprehensively represented, including injury prevention, recovery, exercise progression, older adults, chronic disease and cardiovascular risk. The study did not test personalised exercise prescriptions, recognition of situations in which HIIT may be unsuitable, or correction of misinformation. Reviewer agreement was weaker for overall GQS judgments than for the more structured instruments, and the original reviewer-specific ratings before disagreement resolution were not retained.

The researchers see a supplementary educational role for AI-generated HIIT information, with questions involving symptoms, medical history or individual exercise risk requiring appropriate professional guidance. They call for evaluations that directly check facts, references, safety, comprehension and whether readers can identify and carry out suggested actions. One proposed design would offer a concise plain-language explanation with optional technical detail and verifiable references, an approach whose benefits still need testing.