Human Expert Flexibility Programs Beat Single-Prompt Gemini 2.5 Pro Routine in Trial
Human-expert flexibility programs outperformed a single-prompt Gemini 2.5 Pro routine across every primary outcome in a quasi-randomized trial of 30 male office workers, Nature reported.
Human expert flexibility programs beat a single-prompt Gemini 2.5 Pro routine of static holds, improving sit-and-reach scores by 7.67 cm versus 1.30 cm over four weeks.
Over four weeks, the expert group improved sit-and-reach scores by 7.67 cm versus 1.30 cm for the AI group, while the Oswestry Disability Index dropped significantly only in the human-designed cohort. MyotonPRO stiffness measurements confirmed superior reductions in dominant biceps femoris and non-dominant gastrocnemius for the expert protocol.
The Tech TL;DR:
- Expert dynamic/static protocol beat single-shot AI static routine on flexibility, back disability, and muscle stiffness (p < .001 for sit-and-reach).
- Study design used alternating allocation after a coin toss, introducing selection-bias risk the authors acknowledge.
- Results reflect one specific prompt strategy, not a generalized limit of generative AI for exercise prescription.
Researchers used alternating allocation instead of coin toss
Researchers recruited 30 male office workers and assigned them to a Human Expert Group (n = 15) or an AI Group (n = 15) using a random start followed by alternating allocation. The ethics application had specified simple randomization by coin toss; the deviation creates a predictable sequence that could allow selection bias. Both groups completed 15-minute home sessions three days per week for four weeks.
Expert program outperformed AI on flexibility and disability
Sit-and-reach improved from 14.93 ± 10.67 cm to 22.60 ± 11.49 cm in the expert group and from 16.23 ± 7.91 cm to 17.53 ± 7.60 cm in the AI group. The Time × Group interaction was significant (F(1,28) = 28.26, p < .001, ηp² = 0.502). ODI showed a significant interaction (F(1,28) = 7.49, p = .011, ηp² = 0.211) favoring the expert program. After Benjamini–Hochberg FDR correction across eight interaction tests, sit-and-reach and ODI remained significant.
Muscle stiffness measured by MyotonPRO
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) showed strong Time × Group interactions surviving FDR correction. Dominant erector spinae stiffness was nominally significant before correction (F(1,28) = 4.41, raw p = .045, ηp² = 0.136) but did not survive FDR. Remaining muscle groups showed no significant interactions.
Protocol differences drive the gap
The expert program combined dynamic and static stretching tailored to office-worker biomechanics. The AI program resulted from a single prompt to Gemini 2.5 Pro and consisted solely of static holds. The authors emphasize the findings demonstrate the short-term efficacy of the specific expert-designed dynamic protocol rather than a fundamental limitation of generative AI in exercise prescription.
Nature, the British weekly scientific journal founded in 1869 by Norman Lockyer and Alexander MacMillan, operates editorial offices across London, the United States, continental Europe, and Asia under Springer Nature. It accepts fewer than 8% of submitted papers and, as of 2012, claimed an online readership of about three million unique readers per month.
Registration and next steps
The trial was registered on ClinicalTrials.gov (NCT07242287) on November 17, 2025. Future work should test iterative prompt engineering, multimodal input, and longer intervention periods against expert benchmarks.