Healthcare AI: Separating Fact From Fiction in Physical Therapy and Rehabilitation

Artificial intelligence is now integrated into clinical practice. Physical therapists use AI for documentation, scheduling, patient engagement, remote monitoring, clinical decision support, payment systems, and rehabilitation technologies. For instance, AI-powered scheduling automates appointment reminders and reduces no-shows, while clinical and administrative nudges steer clinicians into better care concordance. 

We’ve passed the point of wondering if AI can fit into rehab therapy. We are at the point of technological evolution where asking “what” AI can do needs to be complemented by “how” and “why”.  We have to look at what evidence supports its use, and stand firm on the point that clinical judgment must remain with professionals.

Distinguishing fact from fiction is essential because AI encompasses a range of technologies. AI-enabled systems include transcription tools, large data processing and categorization, pattern recognition, documentation generation, decision support, and tools that influence access to care. 

What’s Real 

One of the clearest facts about healthcare AI is that it can perform useful tasks. AI systems can process large datasets, recognize patterns, generate text, support clinical analysis, automate portions of administrative workflows, and assist clinicians with information management. More broadly, the FDA's regulatory work recognizes AI’s applications across prevention, diagnosis, treatment, risk assessment, and administrative functions.

Clinical documentation is one of the more common examples in healthcare. Ambient AI scribes can capture conversations between clinicians and patients, transcribe them, and generate structured clinical documentation. A 2025 systematic review of AI-powered documentation systems found evidence of potential improvements in documentation efficiency, while also identifying concerns about note quality, accuracy, and the presence of fabricated or incorrect content. A separate 2025 systematic review of clinical implementation studies found that most studies reported improvement in at least one efficiency measure, but also found variable accuracy and a continuing need for clinician editing. 

This is important. AI scribes do not produce flawless clinical notes. Evidence shows they can assist with documentation and may reduce some burden, but the evidence is limited, and clinicians still encounter errors that require review and correction. A practicing clinician needs to start asking the right questions. It is no longer a question of the presence or absence of the scribe function, but rather of the stability and functionality of this important feature.

Recent mapping studies show rapid growth in AI rehabilitation research. A living systematic review identified 240 studies, with significant activity in neurological and orthopedic rehabilitation. However, more than half lacked a comparator, external validation was uncommon, and explainability was rarely reported. These findings indicate an active research field, but not a mature evidence base.

This is where fact must be separated from marketing. Evidence shows that AI can perform specific tasks and that some AI-enabled rehabilitation technologies can be evaluated clinically. However, there is not yet evidence that AI broadly improves rehabilitation outcomes across all conditions, settings, populations, and care models. That claim requires further evidence.

Where APTA Stands

In 2023, APTA leadership identified artificial intelligence as one of the major issues confronting physical therapy, while discussion within the profession framed AI primarily as an emerging technology with potential to augment practice rather than as an inevitable replacement for clinicians. APTA's 2023 magazine coverage described applications ranging from documentation and patient monitoring to technologies intended to support clinical decision-making and access. 

By 2024, APTA had moved from recognition to formal policy. The House of Delegates adopted a position supporting the ethical development and integration of AI that reduces administrative burden and enhances physical therapist practice, education, and research. The accompanying motion specifically emphasized the need to avoid AI replacing human PT and PTA judgment in critical decision-making processes.

As of 2026, APTA has positioned itself as completely aligned with endorsement of controlled integration and implementation of AI technology among rehab providers and physical therapists. 

What’s Hype

With any form of new technology, there’s a certain amount of hype about what it can do, whether that’s people’s imaginations running wild or companies that overpromise. AI is no exception. 

Assuming AI Will Be Correct

The first form of AI hype is the assumption that technical capability is equivalent to clinical validity. A model can generate a clinically plausible sentence without that sentence being true. A system can identify a statistical pattern without that pattern being clinically meaningful. A prediction can be accurate in one dataset and perform poorly when applied to a different population or clinical environment. FDA guidance specifically recognizes that changes in patient populations, data acquisition, clinical protocols, and real-world environments can affect AI performance. There is no “one size fits all” solution in healthcare technology. The design and architecture of technology platforms have to be built with the foundation of AI and technology ethics. 

Thinking AI Can Replace Clinical Judgment

The second form of hype is the replacement narrative. The existence of an algorithm that can perform one component of clinical work does not establish that the algorithm can replace the clinical reasoning process surrounding that work. APTA's 2024 policy explicitly addresses this issue by supporting AI integration while emphasizing that foundational learning, patient safety, the therapeutic relationship, and human PT/PTA judgment should not be compromised by AI replacement.

Equating More Output With Better Results

The third form of hype is assuming that more documentation means better documentation. Documentation is about communicating the patient’s condition and response to intervention, establishing the clinical reasoning that supports care, supporting continuity among providers, and may carry regulatory and reimbursement consequences. And while generative systems can produce comprehensive-looking notes, length and apparent completeness do not guarantee clinical accuracy. 2025 evidence on AI documentation highlights both potential efficiency gains and concerns about errors, fabricated information, manual correction, and increased note length. 

Any drift or inconsistency in the documentation has to be captured and recorded. Technology that is built on ethical guardrails significantly reduces the probability of system drift and ensures accuracy in the quality of the documentation.  APTA’s practice advisory on AI-enabled ambient scribe technology specifically addresses documentation responsibilities and legal and regulatory considerations rather than treating the technology as an autonomous documentation solution.

Presuming AI Won’t Make “Human” Errors  

There is also a less visible form of hype: treating an AI product’s output as objective simply because it is computational. AI systems inherit properties from their training data, design choices, validation methods, and deployment environment. The FDA identifies bias, uncertainty, limited datasets, and post-market performance monitoring as ongoing areas requiring regulatory attention. This is the one question that consumers need to ask the technology provider: What steps did you take to train, tune, and update the existing model? 

In rehabilitation, this problem can become particularly consequential when systems analyze movement or functional performance. A movement pattern is not automatically pathological because an algorithm labels it differently from another movement pattern. Function exists within context. Age, prior injury, environment, task demands, pain, disability, culture, expectations, and individual goals can all influence how movement is performed. An algorithm that ignores those factors can produce a technically precise measurement with an incomplete clinical meaning. A generic model will always underperform when being operationalized in the rehabilitation setting. 

What’s at Stake in Physical Therapy

Physical therapy occupies an unusual position in health care because its role is often the interaction between a person and their environment. Physical therapists evaluate movement, function, participation, behavior, response to intervention, and changes over time. 

All of this creates a different AI challenge than simply automating transactions.

Documentation integrity becomes central. If AI generates the clinical record, the clinician remains responsible for ensuring that the record accurately represents the encounter. APTA’s 2025 ambient-scribe advisory makes this responsibility explicit and places documentation accountability with the PT or PTA using the technology. 

The Need for Transparency with Patients About AI Use

Patient-facing transparency is equally important. Patients should understand when AI is participating in their care when that information is relevant to their decision-making, privacy, or expectations. For example, a therapist might explain: "Some parts of your care, such as this session's documentation, are assisted by an AI tool that listens and summarizes our conversation to help me with record-keeping. I always review the notes for accuracy, and no information leaves your medical record without your consent. If you have any questions about how this technology is used, please feel free to ask." 

FDA transparency principles identify patients, caregivers, and health professionals as relevant audiences and emphasize communication about intended use, risks, benefits, performance, and other information necessary for safe use.

The Potential for Bias

Algorithmic bias creates another concern. A model developed from a narrow population may not perform similarly across different populations or clinical environments. The FDA's AI program specifically identifies methods for measuring and minimizing bias, evaluating uncertainty, and monitoring performance as areas where additional regulatory science is needed.

The Chance that Clinicians Stop Checking AI’s Work

Clinical reasoning presents a different problem. The risk is not only that an AI system could make a wrong recommendation, but also that clinicians could gradually stop interrogating those recommendations. If technology becomes the primary explanation rather than an input to clinical reasoning, the clinician's role can shift from reasoning to verification. APTA's 2024 policy directly addresses this boundary by opposing the replacement of PT/PTA judgment with AI in critical decision-making. At the same time, clinicians' logic can be reinforced by applied technology. A “clinician in the loop” scenario is a superior way of providing clinical care, as evidenced by multiple studies over the last three years. 

The Risk of AI Creating Incomplete Notes or Improper Denials

Billing creates another layer of both opportunity and risk. AI-generated documentation can speed up the process of creating compliant notes, but it can also make a note appear complete without establishing that every documented service actually occurred. On the other side of the ledger, APTA has taken a position against inappropriate AI use by payers to deny or restrict access to physical therapist services or reduce payment, underscoring that AI can affect both sides of the reimbursement relationship. The same technology, under the correct architecture and design, can be the greatest differentiator in fair and ethical payment capture from the clinician and healthcare system. 

The issue isn’t as simple as whether a rehabilitation AI product is accurate. The profession must consider accountability throughout the entire process: data origin, model development, validation, output presentation, interpretation, action, documentation, and responsibility for errors. This is the challenge of provenance and accountability. Product design, product architecture, and product delivery matters. 

Why the Ethical Foundation Matters

APTA's policy language is instructive because it does not separate innovation from professional responsibility. Its 2024 position supports AI when it reduces administrative burden and enhances practice, education, and research for the benefit of patients, the profession, and society.

The FDA has taken a similar approach at the medical-device level. Its transparency principles emphasize the human-AI team, clear information for users, understanding of intended use, identification of risks and benefits, and the ability to detect errors or declining performance. Its Good Machine Learning Practice principles also emphasize scientifically justified performance measures, representative data, appropriate testing, and lifecycle monitoring.

The principle is straightforward: an AI product should be judged not only by its output but also by whether its use aligns with the clinical environment in which the output is relevant.

This distinction is especially important in physical therapy. A system that saves time but introduces inaccurate clinical information does not necessarily improve care. A triage system that increases access but performs worse in underserved populations does not necessarily improve equity. A movement-analysis system that produces precise metrics without clear clinical meaning does not necessarily improve decision-making.  Evidence and regulatory literature support the need for validation, transparency, bias assessment, and real-world monitoring.

How APTA Has Put Up Guardrails

APTA's progression from 2023 through 2026 reflects this same logic. The profession first recognized AI as a significant issue. In 2024, the House of Delegates established explicit ethical boundaries and opposed inappropriate payer use of AI. In 2025, APTA translated those principles into practice guidance for ambient documentation. In 2026, APTA brought the profession's perspective into federal policy discussions with HHS.

The 2026 HHS comments are particularly relevant because APTA described AI as having the potential to augment physical therapist practice through expanded access, enhanced care delivery models, improved home safety, reduced administrative burden, and improved outcomes. At the same time, APTA's recommendations addressed protection of PT scope of practice and ethical implementation. 

What the Profession Must Do Next

The next step in rehab therapy is not rapid AI adoption but the establishment of clear standards to determine where AI is appropriate.

Build Towards Solving a Problem

Every AI application used in physical therapy should begin with a clearly defined clinical or operational problem. The technology should then be evaluated against the evidence appropriate to that problem. 

• A documentation tool requires evidence about documentation accuracy, completeness, workflow, privacy, and clinician review. 
• A clinical decision-support system requires evidence concerning the relevant clinical outcomes and populations. A movement-analysis system requires appropriate validation against clinically meaningful reference standards. 
• A patient-facing system requires attention to transparency, usability, safety, and the consequences of incorrect information. 

The FDA's AI guidance increasingly organizes evaluation around intended use, performance, risk, transparency, and the total product lifecycle.

Maintain Clinicians’ Authority

The profession also needs to maintain a clear distinction between assistance and authority. AI can summarize information, identify patterns, organize data, automate repetitive tasks, and provide decision support. Those capabilities do not independently establish authority to make clinical decisions. APTA's 2024 policy establishes the professional boundary: AI should enhance practice without replacing PT/PTA judgment in critical clinical decision-making.

Create More Understanding of How AI-Enhanced Systems Work

Clinicians also need to understand the systems they use. That does not mean every physical therapist needs to become a machine-learning engineer. It does mean clinicians need sufficient AI literacy to understand intended use, limitations, validation, data provenance, privacy implications, potential bias, and the difference between a generated output and verified clinical information. Educational resources to support this learning are increasingly available, including materials and courses offered on the APTA website, at professional conferences, and through continuing education providers. APTA's continued development of AI education and practice guidance, and its 2026 request for member input on guidance for artificial and augmented intelligence, reflect this growing professional responsibility.

Require Vendors To Explain What They’ve Built

The market itself has a responsibility. Developers building products for rehabilitation should be able to explain what their systems do, what they do not do, what populations were represented in development and validation, what performance metrics were used, how errors are handled, how model performance is monitored after deployment, and what role the clinician is expected to play. These expectations are consistent with FDA principles for transparency, good machine learning practice, and lifecycle management.

Create Standards to Measure AI’s Performance and Compliance

The profession also needs governance mechanisms capable of evaluating AI after deployment, not simply at procurement. AI systems can interact with changing workflows, patient populations, data sources, and clinical environments. FDA research specifically identifies changes in input data and out-of-distribution cases as factors that can alter clinical utility and safety.

Physical therapy need not choose between innovation and professional integrity. The evidence calls for a more disciplined approach to technology.

AI products in health care must be built on a strong ethical foundation. They must be aligned with clinical purpose, transparent about their capabilities and limitations, and consistent with professional standards. That principle is already visible in the trajectory of APTA policy, from recognition of AI as a major professional issue to formal guardrails, practice guidance, and national advocacy.

The future of AI in physical therapy will be defined not by speed of adoption, but by ethical alignment, clinical integrity, and professional governance.

‍