Research report: AI-generated high-intensity interval training information: a cross-sectional comparative evaluation of quality and readability

When people ask artificial intelligence chatbots about high-intensity interval training, the answers they receive differ sharply in quality and transparency from one platform to another, and none of them come close to meeting a sixth-grade readability benchmark for public-facing health information. That is the central finding of a cross-sectional comparative evaluation published in Frontiers in Public Health, in which researchers systematically compared responses from four major AI systems—ChatGPT (GPT-5.5), Gemini (3.5 Flash), Microsoft Copilot, and Perplexity—using a standardized testing protocol applied to exercise-related health content. The research team, led by Zheng Zhang of Yancheng Institute of Technology together with Xiang-Yan Ruan and Kai-Tuo Zhu, began by identifying what people actually search for.

Using Google Trends set to worldwide interest over the five-year period from August 2021 to August 2026, they collected 29 candidate search terms related to high-intensity interval training. After screening out duplicates, irrelevant entries, navigational terms such as website names, and overly broad queries, 14 queries remained, covering general HIIT terminology, benefits, exercise examples, workouts, beginner-oriented information, and broader concepts such as aerobic exercise, anaerobic exercise, and the named protocol Tabata. The authors note that these broader terms were retained because they commonly appear in explanations of interval training, not because they were treated as synonyms for HIIT.

Each of the 14 queries was submitted exactly once to every chatbot in a fresh conversation, using the exact wording identified through Google Trends, with no follow-up prompts, reformulation, or manual editing. Collection began at 09:00 on August 20, 2026, China Standard Time, using the English-language web interfaces of the four platforms, producing 56 responses in total. The responses were then scored with four established instruments: the DISCERN tool, a 16-item measure of written consumer health information quality with total scores from 16 to 75; the Ensuring Quality Information for Patients (EQIP) instrument, which evaluates content, presentation, and usability as percentages; the Global Quality Scale (GQS), a five-point ordinal measure of overall quality, flow, and usefulness; and the JAMA benchmark criteria, which check four transparency domains—authorship, attribution, disclosure, and currency—each scored 0 or 1 for a maximum of 4.

Perplexity emerged with the highest median scores across all four instruments. Its median DISCERN score of 51.50 fell within the lower range of the good category, and its median EQIP score of 75.00 was the strongest of the group, though still immediately below the threshold for the excellent category. Copilot followed with a median DISCERN of 39.00 and EQIP of 57.50, Gemini recorded 36.00 and 55.00, and ChatGPT had the lowest medians at 35.00 and 47.50.

On the JAMA benchmark, Perplexity was the only platform with a nonzero median, scoring 1.00, while ChatGPT, Copilot, and Gemini each scored 0.00. Median GQS scores were 4.00 for Perplexity, 3.50 for Copilot, and 3.00 for both ChatGPT and Gemini. Because the same 14 queries were submitted to all four chatbots, the observations formed a matched, repeated-measures structure, and the researchers used Friedman tests to assess overall differences, followed by paired Wilcoxon signed-rank tests with Holm correction for the six pairwise comparisons.

All four overall tests were significant: DISCERN yielded a chi-square of 27.98 with Kendall's W of 0.666, EQIP a chi-square of 25.74 with W of 0.613, JAMA a chi-square of 35.40 with W of 0.843, and GQS a chi-square of 21.15 with W of 0.504, all with p values below 0.001. The largest effect size appeared on the JAMA benchmark, indicating substantial between-platform differences in the transparency-related features that instrument captures. In the post-hoc tests, Perplexity differed significantly from each of the other three chatbots on DISCERN, EQIP, and JAMA, while on GQS the Perplexity-Copilot comparison was not significant after Holm correction.

Readability told a more uniformly troubling story. The researchers calculated six established indices—the Automated Readability Index, Flesch Reading Ease Score, Gunning Fog Index, Flesch-Kincaid Grade Level, Coleman-Liau Index, and SMOG—after converting all responses to plain text under a uniform preprocessing procedure that removed URLs, citation markers, and reference lists while preserving substantive wording. A sixth-grade reading level was specified in advance as the benchmark, with grade-level indices required to fall below 6 and the Flesch Reading Ease Score required to reach at least 80.

None of the four chatbots met the benchmark on any measure. Median Flesch-Kincaid Grade Level values ranged from 9.82 for Copilot to 14.40 for ChatGPT, with Gemini at 12.27 and Perplexity at 11.12, meaning even the simplest output remained several grade levels above the sixth-grade target. Median Flesch Reading Ease Scores were 6.47 for ChatGPT, 30.62 for Gemini, 41.02 for Perplexity, and 43.94 for Copilot, all far below the threshold of 80.

Friedman tests confirmed significant between-chatbot differences for all six readability indices. Inter-rater reliability was assessed from the two reviewers' independent ratings before adjudication. Both reviewers held doctoral degrees in sports science, pilot-scored six of the 56 responses to calibrate their interpretation, and were blinded to each other's ratings during formal assessment.

The intraclass correlation coefficient was 0.947 for DISCERN and 0.858 for EQIP, while quadratic-weighted Cohen's kappa was 0.714 for the JAMA benchmark and 0.433 for GQS. The lower agreement on GQS indicates greater variability in global quality judgments, and the authors advise that findings based on GQS be interpreted alongside the more structured DISCERN, EQIP, and JAMA assessments. Disagreements were resolved through discussion and, where necessary, third-reviewer adjudication.

The authors are careful to frame what these findings do and do not mean. The study evaluated the structural quality, usability, transparency, and readability of chatbot-generated text, not its factual accuracy, citation accuracy, hallucination frequency, or clinical safety, none of which were independently assessed. A response can present sources, organize information clearly, and still contain errors, and the JAMA benchmark measures whether authorship, attribution, disclosure, and currency features are present, not whether a cited source exists or accurately supports a claim.

Higher relative scores also do not establish absolute adequacy: even Perplexity's median DISCERN score sat at the lower end of the good category. The results are best interpreted as a comparison of responses generated under a defined protocol in August 2026 rather than a general ranking of the platforms themselves. Several limitations shape the interpretation further.

Each query-platform combination was sampled only once, and because generative systems are stochastic and frequently updated, repeat-query consistency was not assessed. The 14 queries, while drawn from real search behavior, represent a narrow slice of the exercise information landscape, omitting domains such as contraindications, injury prevention, progression, recovery, older adults, and chronic disease. Only English-language responses were analyzed, Google Trends reflects normalized relative search interest rather than absolute search volume, and search-engine terms may not mirror the longer conversational prompts users actually enter into chatbots.

The single-query design did not evaluate multi-turn clarification, personalization, or correction of misinformation, and readability formulas estimate textual complexity rather than actual comprehension. Understandability and actionability were not directly assessed with a validated instrument such as the Patient Education Materials Assessment Tool, so the study cannot determine whether readers could readily identify and carry out recommended actions. Despite these constraints, the study addresses a gap in a literature that has mostly focused on medical rather than exercise content, and it reinforces a consistent pattern from prior chatbot evaluations in other health domains: quality, transparency, readability, understandability, and actionability behave as related but distinct dimensions that do not necessarily improve in parallel.

Notably, the chatbot with the highest quality medians, Perplexity, was not the chatbot with the simplest text, since Copilot had the lowest medians on several grade-level indices. The authors conclude that AI-generated HIIT information may play a supplementary educational role, but that individualized exercise or health decisions should not rely on these outputs alone, and that future evaluations should combine structured quality measures with direct assessment of factual accuracy, citation validity, response variability, and user comprehension. For now, the findings offer a measured benchmark: the best-performing system produced good-quality, better-attributed answers about interval training, yet every system wrote for readers far above the level that public-facing health communication targets.

Subject of Research: Comparative evaluation of quality, transparency, and readability of AI chatbot responses to high-intensity interval training questions

Article Title: AI-generated high-intensity interval training information: a cross-sectional comparative evaluation of quality and readability

Article References: Zhang, Z., Ruan, X.-Y., & Zhu, K.-T. (2026). AI-generated high-intensity interval training information: a cross-sectional comparative evaluation of quality and readability. Frontiers in Public Health, 14, Article 1976676. https://doi.org/10.3389/fpubh.2026.1976676

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1976676

Keywords: AI chatbots, artificial intelligence, high-intensity interval training, health information quality, readability, health literacy, Perplexity, ChatGPT, Gemini, Copilot, DISCERN, large language models