Physical therapists: use AI to catch movement errors, not replace your judgment

Clinicians are already deciding how AI enters their practice

Dr. Claire Morrow, a physical therapist with 13 years of clinical experience and head of clinical consulting at Hinge Health, lays out a framework for responsible AI adoption in physical therapy. The thesis is simple: AI will enter PT clinics regardless. The question is whether clinicians shape that entry or react to it.

Morrow identifies two near-term use cases. First, computer vision technology can monitor patient form during home exercise, offering real-time feedback and progression suggestions without waiting until the next appointment. Second, AI chatbots can reinforce habit-building and pain neuroscience education between sessions, collecting data that flags whether a treatment plan needs modification.

The critical constraint: both tools function as information gatherers and decision-support systems, not decision-makers. The physical therapist remains the primary decision-maker, and the AI output feeds back to the clinician for interpretation and approval.

Morrow also notes a complicating fact: patients are already using large language models to research their conditions. A 2025 study found "high variability and inconsistent accuracy" when comparing widely used LLMs to published clinical practice guidelines for lumbosacral radicular pain. Rather than forbid this behavior, Morrow suggests clinicians invite patients to bring their LLM findings into sessions and use that as a teachable moment to ground care in evidence and context.

Patient adherence and clinician bias are documented problems AI can address—under guardrails

Poor patient adherence is a known primary driver of poor physical therapy outcomes. Patients struggle with uncertainty about exercise form, scheduling friction, or lack of understanding about how a program addresses their condition. Computer vision feedback directly addresses the first barrier. Chatbot support addresses the second and third.

On the clinical side, inter- and intra-rater variability in PT assessments is well-documented. Morrow cites a 2016 study in the Journal of Orthopaedic & Sports Physical Therapy demonstrating anchoring bias: physical therapists measuring wrist passive range of motion gave different results based on historical patient information they received before measurement. AI algorithms, if designed properly, can provide stable baselines and highlight discrepancies that might otherwise reflect clinician bias rather than patient change.

The constraint matters: "Many of these tools cannot yet consistently assess the quality of sources or study design." The clinician's role expands—they must validate the AI input before acting on it. Morrow frames this as synergy, not replacement. The clinician and algorithm together outperform either alone.

Build AI tools with clinicians in your feedback loop from day one

Morrow's strongest practical claim is that tools perceived as opaque, intrusive, or misaligned with clinical goals will be reflexively overridden or rejected. This means vendor-led development without frontline PT input will fail in practice.

Second, recognize that patients are already in the LLM ecosystem. Rather than position AI as intrusive, reframe it as a bridge: patients bring their findings, you coach them toward safer self-education, and you strengthen health literacy. This also gives you an opening to explain your reasoning, building trust and compliance.

Finally, distinguish between tools that extend reach (chatbots, computer vision for home programs) and tools that enhance reasoning (evidence summaries, bias reduction). The first two work best when they reduce friction and gather signal for the clinician. The second works only if the clinician understands and trusts the input.