A recent piece by Beth Rudden, 7 Ways to Spot Someone Who Thinks the AI Understands Them, has been making the rounds and is worth a read. Her main point is one that hits hard: “The chatbot doesn’t have a model of the world. It has a model of how people write about the world.” Most of the time, that distinction does not matter. When it does matter for physiotherapists is in clinical decision-making and citing literature; it’s where things can go wrong.
This is not a reason to back away from AI. It is a reason to step towards it with a plan. At Physiopedia Plus (Plus), we have been wrestling with exactly these questions while developing the Physiopedia AI Assistant (PAI), and a recent article by Beth Rudden, 7 Ways to Spot Someone Who Thinks the AI Understands Them, is a good prompt for sharing where we have landed and what we would suggest to clinicians using AI in their work.
1. The first habit worth building is to verify, rather than ask the AI to verify itself. Rudden describes a familiar pattern: someone asks a Large Language Model (LLM) like Claude or ChatGPT to check its own answers, and the grader and the student turn out to be the same machine. For physiotherapists using AI to find research or check guidelines, this matters in a very practical way. If a general LLM gives you a citation for a systematic review on patellofemoral pain, asking that same LLM whether the citation is real does not verify anything. Verification means opening and critically reading the journal article. It is the same standard we have always applied to evidence, just applied to a new source.
2. The second habit is to remember that the apology is not the correction. When an LLM responds to pushback with “you are right, I should have been more careful”, those words are shaped like remorse, but nothing underneath has actually learned anything. So the corrected answer deserves the same scrutiny as the first one. If a tool gave you an incorrect dose, technique or contraindication, a confident-sounding revision is not evidence that the new version is right.
3. The third habit is to be wary of role prompts. Telling a general LLM “you are a board-certified neurological physiotherapist with twenty years of experience” does not summon that expertise. It picks a writing style. The output will sound like something an experienced clinician might write, which can genuinely help with brainstorming or drafting, but it does not carry the clinical reasoning, accountability or experience that a real practitioner brings. By telling the LLM who it is, you are telling it to put on a costume. The costume is useful when you remember it is a costume.
This is also where the design of an AI tool starts to matter. When we built PAI we looked at how it could respond differently to a general LLM, in ways that directly address these concerns. It draws on Physiopedia’s curated, peer-reviewed content and the structured material within Plus courses, rather than the open internet. So when PAI guides a learner through a clinical reasoning activity, it is working from resources written and reviewed by physiotherapists, under editorial oversight, rather than predicting what plausible-sounding text might come next from somewhere on the web.
Just as importantly, PAI shows its working. Responses link back to the specific Physiopedia pages and Plus courses the content originated from, so clinicians can follow the trail to the source, check the underlying material, and decide for themselves whether they agree with how it has been applied. This is the verification habit built into the tool itself, rather than left for the user to remember. It does not make PAI a substitute for clinical judgment, and it is not meant to. It does mean clinicians can engage with PAI the way they would with a well-referenced clinical resource, using it to think with rather than to think for them.
The practical takeaway is simple. When using general AI tools, treat the output as a starting point and not an answer. Check the references against the original sources. Confirm techniques against current clinical guidelines. Hold any clinical claim to the same evidence standard you would apply to a colleague’s recommendation in the corridor.
There is one more idea from Rudden’s piece worth holding onto. She writes about being polite to AI, saying please and thank you, and makes the point that manners are not really about the LLM or AI tool, even though it improves the output. They are about the practitioner. The LLM does not understand you, but how you speak shapes who you become through the conversation.
For clinicians, this lands somewhere familiar. The way we speak to colleagues, to patients, to ourselves in our own clinical notes builds the practitioner we are. AI is just another room we walk into. The question is who walks out.
This article was written by one of our course instructors...
Tarina van der Stockt
PT (BphysT), DPT
Education Director Physiopedia Plus