How AI in rehabilitation helps plan recovery, but still needs a doctor's eye
A review of nine studies found that generative AI can draft accurate physical recovery plans, matching expert opinion up to 98.75% after review. However, models still invent false medical details, meaning clinicians must oversee every plan.
Reading level
The full story with proper science words explained.
A helpful helper that makes things up
Think about healing after heart trouble or a broken bone. A physical therapist spends hours writing a custom workout to help you heal. Now, scientists are testing whether AI in rehabilitation can draft these therapy plans faster without risking patient safety.
Testing the software against real doctors
A systematic review—a detailed study that collects and checks all existing evidence—searched three major medical databases: PubMed, Web of Science, and Scopus. Across all published papers up to February 2026, only nine studies rigorously tested generative large language models (LLMs). These are AI systems trained on text to produce human-like writing.
The scientists checked how well AI plans matched advice from human experts and official clinical guidelines, which are rulebooks for doctors. Two people checked each paper using JBI tools, which are quality checklists from the Joanna Briggs Institute. A third person settled any split votes.
The test results showed clear promise, but also real risks. The models followed care rules well for muscles, bones, and heart care. When an expert panel reviewed and cleaned up the AI drafts, agreement with specialist doctors reached 98.75%. Yet without human help, the bot varied widely and often suffered from hallucinations—a problem where an AI states made-up facts as the truth.
Why human eyes are still needed
An LLM works like an eager intern. It writes notes in seconds, but it sometimes makes things up with total confidence. A hospital cannot leave an intern alone on the ward. In the same way, doctors cannot let software plan care on its own. Human clinicians must check every step to keep patients safe.
What scientists still need to learn
Because only nine studies met the rules for this review, our evidence base is still quite small. Scientists have not yet tested how well these AI plans help real patients in actual clinics. Hospitals also lack clear rules for bringing these smart tools safely into daily care.
Science words
- Systematic review
- A rigorous scientific study that collects, checks, and combines all existing evidence on a specific question.
- Large language models (LLMs)
- A type of generative artificial intelligence trained on massive text datasets to create human-like responses and new content.
- Clinical guidelines
- Systematically developed statements and rules to assist healthcare practitioners with decisions about appropriate care.
- JBI tools
- Standardised critical appraisal checklists developed by the Joanna Briggs Institute to evaluate study quality and risk of bias.
- Hallucinations
- Instances where an AI model generates plausible-sounding but factually incorrect, ungrounded, or fabricated information.
Check it yourself
This story is based on a real research paper in American Journal of Physical Medicine & Rehabilitation by Demeco, Marotta, Racinelli et al. We write with AI help and check it against the paper, but the original is the final word.