Human expert-generated flexibility program versus single-attempt AI (Gemini 2.5 Pro)-generated flexibility program for office workers: a quasi-randomized controlled trial
Sedentary office environments contribute to flexibility loss and musculoskeletal disorders. While Artificial Intelligence (AI) offers scalable rehabilitation, its clinical reasoning and biomechanical safety remain scrutinized in academic literature. This study compared the efficacy of 15-minute home-based flexibility exercises designed by human experts versus those generated by AI (Gemini 2.5 Pro) on flexibility, lower back health, and muscle stiffness. In this quasi-randomized controlled trial, 30 male participants were allocated to a Human Expert Group ( n = 15) or an AI Group ( n = 15) using a random start (single coin toss) followed by alternating allocation, which differed from the wording of the ethics-approved application (simple random method, coin toss) and carries a risk of selection bias. Interventions lasted 4 weeks (3 days/week). Assessments included Sit-and-Reach tests, Oswestry Disability Index (ODI), and MyotonPRO stiffness measurements. Initially, the groups exhibited similar characteristics ( p > .05). In the Sit-and-Reach test, a measure of flexibility, the average score for the human expert group increased from 14.93 ± 10.67 cm to 22.60 ± 11.49 cm, while in the artificial intelligence group, this value reached only from 16.23 ± 7.91 to 17.53 ± 7.60 cm; a statistically significant Time × Group interaction with a large effect size was detected between the groups (F(1,28) = 28.26, p < .001, ηp² = 0.502). A significant Time × Group interaction was also found for the Oswestry Disability Index (ODI) (F(1,28) = 7.49, p = .011, ηp² = 0.211). Following Benjamini–Hochberg False Discovery Rate (BH-FDR) multiplicity adjustment across the eight interaction tests, statistically significant Time × Group interactions robustly survived for Sit-and-Reach flexibility ( p < .001), Oswestry Disability Index ( p = .011), dominant biceps femoris stiffness (F(1,28) = 11.01, p = .003, ηp² = 0.282), and non-dominant gastrocnemius stiffness (F(1,28) = 11.16, p = .002, ηp² = 0.285) in favor of the human expert group. In contrast, the Time × Group interaction for dominant erector spinae stiffness was nominally significant prior to adjustment only (F(1,28) = 4.41, raw p = .045, ηp² = 0.136) and did not survive FDR correction. No significant interactions were observed for the remaining muscle groups ( p > .05). Based on the observed Time × Group interaction effects, human expert-generated programs containing dynamic and static flexibility exercises yielded significantly greater improvement over time in flexibility and lower-back health than the single-shot static stretching program generated by Gemini 2.5 Pro. These findings highlight the short-term efficacy of the specific expert-designed dynamic protocol rather than demonstrating a generalized limitation of generative AI in exercise prescription. Trial registration : The study was registered on Clinicaltrials.gov under the number NCT07242287. (Date of Registration 11/17/2025)
Authors
- Esedullah Akaras (ORCID: https://orcid.org/0000-0002-0305-4632)
- Mehmet Sönmez (ORCID: https://orcid.org/0000-0002-3617-9087)
Institutions
- Erzurum Technical University (TR)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1038/s41598-026-81629-0
- Primary Topic
- Musculoskeletal pain and rehabilitation
- Type
- article
- Field-Weighted Citation Impact
- 0.00