Will AI Replace Physical Therapists?
Among the lowest AI coverage of any occupation in the real-usage data: treatment is hands-on, individual and physical. AI helps with notes and exercise libraries; it does not put hands on a patient. An ageing population drives 11% projected growth.
The exposure evidence — three sources, side by side
There is no official "AI risk score" for any job, and any site that gives you one number without a source is making it up. Here are the three serious research measures that exist for this occupation, each with what it actually counts:
| Research source | Exposure figure | What it actually measures |
|---|---|---|
| Anthropic Economic Index — observed exposure | 2% of tasks | Share of this occupation's tasks where AI use was actually observed in real Claude conversation data (Anthropic, "Labor market impacts of AI", March 2026 dataset). This is measured use, not a prediction. |
| GPTs are GPTs (OpenAI / University of Pennsylvania) — GPT-4 rating, β measure | 19% of tasks | Share of tasks rated exposed to large language models, counting direct exposure in full and assisted exposure at half weight (the paper's β measure). Human expert raters scored the same occupation at 34%. |
| Microsoft "Working with AI" — AI applicability score | 0.17 / 1.00 | A 0–1 score built from 200,000 real Bing Copilot conversations, weighting how often AI was used for this occupation's tasks and how well it completed them. It is a score, not a percentage; the median occupation scores about 0.14. |
The three sources measure different things — observed use, rated capability and Copilot-based applicability — so they rarely agree exactly. Where they disagree, that disagreement itself is information: it usually means a job is being changed by specialist tools, or in ways general chatbots do not show. All three agree on the framing: exposure is not replacement.
The work itself: what AI can help with, and what stays human
These are this occupation's real tasks, from the US Department of Labor's O*NET database (version 30.0). The split is our editorial analysis — the tasks on the left are the routine, screen-based ones current AI tools demonstrably assist with; the ones on the right involve physical presence, licensed responsibility, or judgement about people.
Where AI can already help
- Recording prognosis, treatment, response and progress in the patient chart
- Documenting goals, anticipated progress and re-evaluation plans
- Providing educational information and home-programme instructions
What stays human
- Examining patients and evaluating movement, strength and function in person
- Administering manual therapy, exercise and physical agents by hand
- Adjusting treatment live, based on how this patient responds today
- Obtaining informed consent and explaining risks, benefits and alternatives
Pay, outlook and entry requirements (BLS)
| Median annual pay, 2024 | $101,020 |
|---|---|
| Employment change, 2024–34 | BLS projects employment to grow by 11% from 2024 to 2034 (267,200 → 296,400 jobs, a change of +29,300). |
| Typical entry-level education | Doctoral or professional degree |
Basis: BLS Occupational Outlook Handbook, 2024–25 edition profile (2024 pay, 2024–34 outlook). A median is the midpoint of what all workers in the occupation earn — it is not a starting wage, and it is not a forecast.
What to learn next
A roadmap only helps if it starts from the skills this job already builds. These three moves do — each is chosen because the destination work is growing, harder to automate, or carries a credential software cannot hold:
- Telehealth and remote-monitoring programmes → a stronger position in the same field. Hybrid care — in-person assessment plus monitored home exercise — is expanding; therapists who design and run those programmes scale their impact without losing the human core.
- A clinical specialism (sports, neuro, geriatrics, pain) → a stronger position in the same field. Specialist caseloads are more complex, less protocol-driven and better paid — and ageing demographics push hardest on geriatric expertise.
- Clinic ownership / service leadership → a stronger position in the same field. Demand growth rewards those who build services, supervise teams and own the patient relationship end to end.
AI tools in this job today
Documentation assistants, exercise-video libraries and wearable data summarise the admin around therapy. The therapy itself — assessment by hand and eye, manual treatment, motivation — is embodied work that software supports but cannot perform.
Frequently asked questions
Will AI replace physical therapists?
Not on any schedule the research supports. The strongest exposure figure for this occupation is low — about 19% of its tasks are within AI's reach — but exposure measures tasks, not whole jobs. BLS projects this occupation growing 11% between 2024 and 2034, and Anthropic's own labour-market research found no systematic rise in unemployment among workers in highly exposed occupations. The realistic outcome is a changed job, not an abolished one.
Which parts of a job are most exposed?
The tasks listed under “Where AI can already help” above — mostly the routine, text- and data-handling parts of the role. They come from this occupation's real O*NET task list, and they are the parts to stop competing on and start supervising.
How much do physical therapists earn?
The BLS median annual pay was $101,020 in 2024 — half earned more, half earned less. It is a national median, not a starting salary or a promise; pay varies widely by state, employer and experience.
What should I learn next?
The roadmap above is built from this occupation's actual skills: Telehealth and remote-monitoring programmes; A clinical specialism (sports, neuro, geriatrics, pain); Clinic ownership / service leadership. Each points at work that is either growing, harder to automate, or both — with the BLS outlook for the target occupation stated next to it.
Where do these numbers come from?
Exposure figures come from three published research datasets — Anthropic's Economic Index, the OpenAI/University of Pennsylvania GPTs-are-GPTs study, and Microsoft's Working with AI paper. Pay, outlook and education come from the US Bureau of Labor Statistics (Employment Projections and Occupational Outlook Handbook, 2024–34 cycle). Every figure on this page is dated and sourced in the Sources section below, and our full method is on the Methodology page.
Sources
- Anthropic, "Labor market impacts of AI" (Massenkoff & McCrory, March 2026) and the Economic Index job-exposure dataset (observed exposure by occupation). Accessed 11 November 2026.
- Eloundou, Manning, Mishkin & Rock, "GPTs are GPTs" (OpenAI / University of Pennsylvania, 2023), occupation-level exposure data (β measure), official dataset. Accessed 11 November 2026.
- Microsoft Research, "Working with AI: Measuring the Applicability of Generative AI to Occupations" (2025), AI applicability scores (CC BY 4.0). Accessed 11 November 2026.
- US Bureau of Labor Statistics, Employment Projections Table 1.2 and Occupational Outlook Handbook (2024–34 cycle): pay, outlook, education. Public domain. Accessed 11 November 2026.
- US Department of Labor, O*NET Database 30.0 (CC BY 4.0): occupation tasks and skills. Accessed 11 November 2026.
Data checked November 2026. Figures are national US figures. Exposure scores describe tasks, not people, and none of the sources above predicts that this occupation will be eliminated.