Human Expert Programs Beat AI in Flexibility Trial for Office Workers

Human experts outperformed Gemini 2.5 Pro in a four-week trial, with participants showing significantly greater improvements in flexibility and lower-back health after following expert-designed routines.

The study tracked 30 male participants who were split into two groups to test the efficacy of 15-minute home-based flexibility exercise programs. While both groups engaged in three sessions per week, those following the human-expert protocol achieved a Sit-and-Reach test increase from 14.93 ± 10.67 cm to 22.60 ± 11.49 cm. In comparison, the group using the AI-generated routine saw their scores move from 16.23 ± 7.91 cm to only 17.53 ± 7.60 cm.

Researchers utilized the Oswestry Disability Index (ODI) and MyotonPRO technology to quantify physical changes. The findings revealed a statistically significant difference in how the two groups responded over the four-week period, with the expert-led group showing superior outcomes in lower-back health.

After applying Benjamini–Hochberg False Discovery Rate adjustments, the researchers confirmed that the human-expert group maintained strong advantages in flexibility, ODI scores, and the stiffness levels of specific muscle groups, including the dominant biceps femoris and non-dominant gastrocnemius.

The study said that the AI-generated program relied on single-shot static stretching, whereas the human-designed protocol incorporated a mix of dynamic and static movements. This difference in design strategy accounts for the variation in results, rather than showing a permanent inability for generative AI to assist in rehabilitation.

The authors emphasized that the results highlight the efficacy of the specific dynamic protocol used by experts. The study, registered on Clinicaltrials.gov under NCT07242287, remains a point of reference for the ongoing scrutiny of AI clinical reasoning and biomechanical safety in digital health.

Why did the human-designed program perform better?
The human-expert program utilized a combination of dynamic and static flexibility exercises, whereas the Gemini 2.5 Pro model generated only a single-shot static stretching routine.

Did the AI group see any improvement?
Yes, both groups showed changes, but the improvements in the human-expert group were significantly larger, with a large effect size detected in the Sit-and-Reach test (F(1,28) = 28.26, p < .001, ηp² = 0.502).

Does this mean AI should not be used for exercise?
The study concludes that these findings highlight the short-term efficacy of the expert-designed dynamic protocol rather than demonstrating a generalized limitation of generative AI in exercise prescription.