Somewhere between the twinge and the appointment, most people now open a chatbot. You describe the pain, it replies in seconds with a confident-sounding name for it, three exercises and a note about seeing a professional if things do not improve. It will sound plausible. The question worth asking is which parts of it you can safely act on.
The honest answer is that a large language model is genuinely useful for some of what you want from a clinician and structurally incapable of the rest. Knowing which is which saves you the two most common outcomes we see: weeks of the wrong exercise, and a delayed appointment for something that needed one quickly.
Why so many people ask a chatbot first
It is free, it is available at eleven at night when the shoulder wakes you up, and it does not make you explain yourself twice. It also answers the question people are too embarrassed to ask a professional: is this actually bad enough to bother someone about?
This has become common enough that the profession has started writing about how to handle it. A 2026 editorial in Physiotherapy, the Chartered Society of Physiotherapy's journal, argues that clinicians should engage with the AI advice patients bring into the room rather than dismiss it, because people are already using these tools between appointments and sometimes instead of them. Physiotherapists themselves are cautious about it: in a cross-sectional survey of 415 Italian physiotherapists published in Frontiers in Digital Health in 2025, 84.4% named patient self-diagnosis as a risk and 72.1% named the spread of false information, even though most were positive about the technology's future.
So the concern is specific, and it sits downstream of the writing. The model handles anatomy competently; the trouble starts when a person applies a general answer to a body it has never examined.
What AI actually does well
Used as a reference rather than a clinician, these tools are strong at four things:
- Translating jargon. If a letter or scan report says "mild subacromial impingement with acromioclavicular degenerative change", a chatbot will turn that into plain English quickly and usually accurately.
- Explaining a mechanism. Why tendons hurt more the morning after, why nerve pain travels, why a joint feels stiff for the first ten minutes of the day. General physiology is well-represented in the training data and stable over time.
- Preparing you for an appointment. Ask it what a physiotherapist will want to know about your problem and you will arrive with a much better history than most people manage.
- Reassurance about ordinary things. Muscle soreness two days after unusual exercise, a joint that clicks without pain, bruising that changes colour. These are common, benign, and well-documented.
Notice that all four are about understanding. None of them require the model to know anything about your particular body.
The part it cannot reach: what an examination actually finds
A diagnosis in musculoskeletal care is rarely made from a description. It is made by comparing sides, loading a structure and seeing what provokes the pain, testing strength against resistance, checking how a joint moves when you are not the one moving it, and watching you do the thing that hurts.
Take a simple example. "Pain on the outside of the hip when I lie on that side" is a description that fits gluteal tendinopathy, a lumbar spine referral, bursitis and a hip joint problem. Those four have different treatments, and two of them get worse with the stretches most commonly recommended for the other two. A clinician separates them in about four minutes with resisted tests and palpation. A chatbot cannot separate them at all, so it does the reasonable thing and gives you the most common answer, which is right often enough to be dangerous.
The same applies to the two questions people actually want answered: how bad is it, and how long will it take? Both depend on your history, your loading tolerance, what you need to get back to, and how the tissue responds over the first fortnight of treatment. None of that is in your prompt.
A useful test: if the answer would change depending on what a clinician found when they pressed on it, moved it, or watched you walk, then the chatbot is guessing. That is most of diagnosis.
The warning signs no chatbot should be the last word on
Some symptoms need same-day assessment regardless of how reassuring an AI answer sounds. Chatbots do often flag these, but they flag them at the bottom of a long answer, after the exercises, in the paragraph most people skim. Take them out of that order and treat them as the first question, not the last.
- Numbness around the saddle area, loss of bladder or bowel control, or numbness and weakness in both legs. With back pain, this combination can indicate cauda equina syndrome, which the NHS treats as an emergency. Go to A&E.
- An inability to bear weight after an injury, or a joint that looks visibly out of shape.
- Sudden severe calf pain with swelling, warmth and redness, particularly after surgery, immobility or a long flight. This needs ruling out on the day, not stretching.
- Unexplained weight loss, night pain that wakes you every night, fever, or a history of cancer alongside new persistent pain.
- Pain following a significant fall, particularly over 50 or with osteoporosis.
For anything urgent but not an emergency, NHS 111 is the right first call. Nothing in this article, and nothing a chatbot tells you, replaces that.
How to use AI without it costing you time
The failure mode is asking, getting a confident answer, and then spending six weeks testing it on yourself. Three habits prevent that:
- Set a review date before you start. Decide up front how long you will give self-management, and book something if you reach it. Two to three weeks with no clear direction of travel is a sensible ceiling for most new musculoskeletal pain.
- Ask it what else this could be. Prompt for the differential rather than the diagnosis: "list the four most likely causes of this pattern and what would distinguish them on examination". You get a far more honest answer, and it tells you what to ask a clinician.
- Never let it override a worsening symptom. If pain is spreading, waking you, or producing new numbness or weakness, the exercise plan is no longer the relevant question.
It is also worth knowing that these tools rarely tell you when they are uncertain. A model that has never seen your hip produces an answer in exactly the same confident register as one describing a well-documented fracture. That confidence is a property of how the software writes, not of how much it knows about you.
What to bring to an appointment instead
If you have already done the research, put it to work. The people who get the most out of a first physiotherapy session tend to arrive with four things: when the pain started and what you were doing, what makes it clearly better and clearly worse, what you have already tried and for how long, and what you specifically want to get back to. That last one shapes the plan more than the diagnosis does.
Bring the AI answer too, if you have one. A good clinician would rather correct a specific belief than work around an unspoken one, and if it turns out the chatbot was right, you have lost nothing. If you want a sense of what the session itself involves, our network's clinic site walks through what happens at a first physiotherapy appointment, and there is a plain guide to the routes to seeing a physiotherapist in the UK if you are weighing up NHS against private.
How this works at PhysioNearby
The question we are asked most often is a version of the one above: "is this worth getting looked at, or am I overreacting?" It is almost always asked by someone who has already searched, already read three contradictory answers, and would quite like one person to just tell them.
The reason that question is hard to answer online is that the useful part is a comparison. Pain on its own means very little; pain that behaves differently from the other side, or that changes in a specific way under load, means quite a lot. That is also why the waiting question matters so much locally. The East Sussex MSK Community Partnership publishes its current physiotherapy waits by location, and on the June 2026 figures the averages ran from under seven weeks in Hastings to around ten in Rother, with the longest individual waits considerably beyond that. Several weeks of guessing from search results is a genuinely different experience from several weeks with a plan and a date to review it.
PhysioNearby matches you with a vetted local physiotherapist across Sussex, from Brighton and Worthing to Uckfield and Bognor Regis, so that the assessment part happens early even when the treatment runs on afterwards. Start a request and a clinician reads it before anyone is matched.
Common questions
Is it dangerous to ask AI about my symptoms?
The risk sits in acting on a diagnosis it cannot make, particularly when the plan involves loading a structure that turns out to be something else, or when it delays an assessment for a symptom that needed one quickly. Asking the question itself costs you nothing.
Can AI tell me which exercises to do?
It can give you a reasonable general set for a named condition. What it cannot do is confirm you have that condition, or set the dose. Dose is where most rehab fails: the right exercise at the wrong load or frequency will either do nothing or flare the problem.
Will a physiotherapist mind if I mention what ChatGPT said?
Most would rather hear it. It tells them what you are worried about, which is often more useful than the symptom description itself.
Are the sources it cites reliable?
Sometimes, and it can also produce references that do not exist. If a claim matters to your decision, check it against a named source you can open yourself, such as NHS, NICE or a published review.
Use the chatbot for the explanation and the questions. Use a clinician for the examination and the dose.