Why this specialty keeps a person in the room
Physiatrists treat people whose bodies have stopped doing what they used to do. A stroke, a spinal cord injury, a knee replacement, chronic back pain. The work starts with a physical exam: watching someone stand, testing strength and tone, feeling a joint, checking how far a limb moves before it hurts. Software can read the notes that come out of that exam. It cannot do the exam.
A second block of the job is procedural. Electrodiagnostic testing, joint and spine injections, spasticity management, tone-reducing treatments. These need steady hands, a needle in the right place and a judgment call made while the patient is in front of you. Our robotics read puts about a third of this job’s task time in the physical category, at a difficulty tier that assumes a dexterous humanoid. That hardware is not sitting in rehab units.
The third block is the part people underrate: setting goals. A physiatrist decides what recovery should aim for, then holds a team of therapists, nurses and case managers to it, and tells a family what is realistic. Asking whether AI will replace rehabilitation physicians mostly comes down to whether a model can own that call. Nothing in our evidence says it can.
What AI does, what it assists, what stays with people
Start with the share AI can run on its own: 0%. That sits in the paperwork layer of the specialty. Drafting clinic notes and discharge summaries from a recorded visit, coding and billing support, pulling prior imaging and therapy records into one place, tracking functional outcome scores across an admission. None of it touches the patient, and all of it is where physician hours quietly go.
Next, the assisted share: 42%. Here the model prepares work a physician signs. Flagging patterns in electrodiagnostic or imaging data for review, suggesting a starting rehab protocol from a diagnosis, summarizing drug interactions before a spasticity plan, reading remote sensor data between visits. The physiatrist still decides, and still carries the liability for the decision.
The remainder is the part that stays with a person: 58%. Hands-on examination, injections and procedures, goal-setting conversations with patients and families, and leading the rehab team through a case that is not going to plan. Overall coverage, our estimate of how much task time AI can handle today, reads 19 out of 100. The coverage method explains how that split is built.
What the evidence shows, and what it does not
Our quality-parity grade for this job is D. A D grade means one thing: nobody has run a credible head-to-head test of AI against physiatrists on their own work. There are studies of models reading images and of AI tools in therapy settings, but none of them measures what a rehabilitation physician does across a case.
What would settle it is specific. A prospective study comparing physician-set rehabilitation plans with model-generated plans, judged on function at discharge and at six months. An audited comparison on electrodiagnostic interpretation against physician readings. An error and harm rate for AI-suggested injection and spasticity decisions. Until work like that exists, we publish no parity number for this job rather than guess one. The quality-parity method sets out how the grades are assigned.
The market context is steadier than the discussion around it. The Bureau of Labor Statistics counts 342,720 people in this occupational group and a median wage of $265,930 (BLS, 2025), with projected employment change of 3.4% through 2035. That is slow growth, not contraction.
When this could change
Most likely after 2042 (8 in 10 of our scenarios). The replacement-year method explains what that window is measuring and how wide it is meant to be.
Two things could pull it earlier. The cheap end of AI documentation and triage is already in hospitals, and once it handles the record-keeping layer, employers reprice the rest of the job. Remote sensors and home exercise monitoring could also move follow-up visits out of the clinic, which thins the number of physician touchpoints per patient.
Two things hold it back. The procedural core needs hands that can place a needle under ultrasound, and that hardware is nowhere near routine clinical use. And medicine gates practice through licensure, credentialing and liability. A model that cannot be sued, credentialed or supervised does not get to sign the plan, however good its draft is.
What to do: treat AI scribing and summarization as something to adopt early and audit hard, rather than something to wait out.
How to stay needed in physiatry
Lean into the parts of the job in the needs-a-human column. First, procedures: electrodiagnostics, image-guided injections and spasticity management keep you in work that no software performs. Second, the physical exam and the functional assessment that follows it. Third, the goal-setting and team leadership that turn a diagnosis into a plan other clinicians can run.
Two skills are worth adding. One is reading machine output critically: knowing when a flagged finding or a suggested protocol is wrong, and being able to say why in the chart. The other is rehabilitation outcome measurement, because the physician who can show function gained per dollar is the one who shapes how AI tools get deployed, instead of being measured by them.
Close neighbors are worth a look too. Compare this role with sports medicine physicians, neurologists and physical therapists, or put any two of them side by side with the job comparison tool. The wider picture sits in the diagnosing and treating practitioners family and the healthcare sector page, and this job’s position among jobs that mostly need a person is listed there.
The headline figure above, 78 out of 100 (higher is safer), comes from open data and a published method. You can read how the scoring works, or check where your own role lands in the full job rankings.