Generative artificial intelligence in exercise and physical therapy: a systematic review of emerging evidence and clinical implications
Yi ChungBackground/Aims
With generative artificial intelligence tools increasingly being incorporated into clinical and community exercise settings, a synthesised appraisal of the available evidence is urgently needed to inform safe and effective practice. The aim of this study was to systematically review evidence on the use of generative artificial intelligence applications in exercise prescription, movement analysis and rehabilitation within physical therapy.
Methods
PubMed, Scopus, and IEEE Xplore were searched for English-language studies published between January 2015 and July 2025. Peer-reviewed studies using generative artificial intelligence models (large language models, generative adversarial networks, variational autoencoders or transformer-based generative architectures) applied to exercise or rehabilitation contexts were eligible. Two independent reviewers screened records; 12 studies ultimately met inclusion criteria after full-text review of 50 reports identified from 318 database records. Extracted data included study design, artificial intelligence type, application domain, sample size, outcomes, and reported benefits and limitations. Risk of bias was formally assessed using Revised Cochrane Risk of Bias Tool for Randomised Trials (for randomised controlled trials), Risk of bias in non-randomised studies of interventions (for non-randomised studies), and the METRICS checklist (for artificial intelligence model evaluation studies).
Results
Formal meta-analysis was not conducted. One randomised controlled trial and one quasi-experimental study were identified but are presented as a descriptive narrative comparison only, as pooling was determined to be methodologically unjustifiable owing to fundamentally incomparable comparators, directionally opposite outcome metrics, and extreme statistical heterogeneity. Across the remaining 10 included studies, generative artificial intelligence demonstrated potential for individualised exercise plans, motion analysis using synthetic data, and virtual coaching.
Conclusions
Generative artificial intelligence is a promising adjunct in physical therapy, supporting personalised exercise prescription and movement analysis, but requires human oversight to ensure safety and effectiveness. Current evidence emphasises the use of artificial intelligence as supportive rather than standalone. Future research should prioritise adaptive real-time systems, large-scale trials with sex-stratified reporting, and clinical ethical frameworks.
Implications for practice
Although AI exercise programming shows promise as a supportive tool, it is inferior to expert-designed programmes in direct comparisons, lacks specificity and personalisation, and introduces equity concerns. The best outcomes emerge from AI–human collaboration – AI-generated plans refined by expert clinicians.