Can AI on Your Clinic Website Avoid Giving Medical Advice? Where the Line Sits
Key Takeaways
- The line sits between information about your clinic and advice about a person. Services, fees, first visits and booking are information. "What is wrong with me", "should I stop my medication" and "will this treatment work for me" are advice.
- Patients already ask AI health questions. About one in ten Australian adults asked ChatGPT a health question in the first half of 2024, and 61% of them asked at least one question that would usually need a clinician.
- A clinic assistant should answer practice facts only from what you have published, and refuse diagnosis, treatment and suitability calls, plus business facts your site does not state, such as funders, age limits or prices.
- In QuackChat, the rule against personalised medical advice is the last instruction in every prompt, and your own custom instructions cannot remove it.
- You can audit the line yourself: ask the hard questions on the dashboard's Test page before launch, then read real transcripts and the sources behind each answer in Logs.
An AI assistant on your clinic website should give information about your practice and never advice about a patient: it can explain what you treat, what a first visit involves, what it costs and how to book, but it should not diagnose, recommend treatment or judge whether a service suits someone. That line protects patients and your registration. This guide shows where it sits, what a well-built assistant refuses, how QuackChat holds the line even against its own settings, and how to check it yourself.
Contents
- Why are clinic owners right to be wary?
- Where is the line between information and advice?
- Where does a grounded assistant get its answers?
- What should it refuse to answer?
- Can your own settings switch the safety rule off?
- What happens when a patient describes symptoms?
- How do you audit where your assistant draws the line?
- Frequently Asked Questions
- References
Why are clinic owners right to be wary?
Because patients will ask a chatbot medical questions whether or not it should answer them. A nationally representative survey of 2,034 Australian adults, published in the Medical Journal of Australia, estimated that 9.9% had used ChatGPT for health information in the first half of 2024. Of those users, 61% had asked at least one higher-risk question, meaning one about taking action that would typically need clinical advice, and 37% had asked what their symptoms meant [1].
General-purpose AI is also weaker at this than its test scores suggest. In a randomised study of 1,298 UK participants published in Nature Medicine, language models working alone identified the relevant condition in 94.9% of cases. When members of the public used the same models, they identified relevant conditions in fewer than 34.5% of cases, no better than people using their usual resources [2].
A chatbot on your website speaks in your clinic's name. Whatever it says about a patient's knee, the patient hears from you.
Where is the line between information and advice?
Information is about your practice and applies to anyone. Advice is about one person and needs professional judgement.
| Information (fine) | Advice (not fine) |
|---|---|
| "We treat sports injuries and post-surgical rehab." | "That sounds like a meniscus tear." |
| "A first appointment takes 45 minutes and includes an assessment." | "You should try dry needling for that." |
| "An initial consult is $110." (if your fees page says so) | "You will only need three sessions." |
| "Here are Thursday's openings with our physios." | "It is safe to keep running on it." |
The regulators draw a similar line from their side. Ahpra's guidance on AI says the practitioner "remains responsible for delivering safe and quality care" when AI is involved [3]. The TGA lists chatbots that "suggest, deliver or monitor treatment" among its examples of AI medical devices, and says that when a model built for general information is used to give diagnoses and treatment strategies, that use likely meets the definition of a medical device [4].
None of this is legal advice. It does explain why the safest assistant stays firmly in the left-hand column.
Where does a grounded assistant get its answers?
From your own website, and nowhere else, for anything about your practice. QuackChat reads your pages and answers from them. If your site says you see children from age eight, it can say so. If your site is silent on age limits, it says it does not have that detail and points the patient to your phone number or booking link.
There is one deliberate exception. Asked a question about the profession in general, such as whether physiotherapists commonly treat teenagers, it can answer what is usual for the profession, framed that way, and never as a statement about your clinic. It then suggests the team can confirm.
That exception has a hard boundary. It never covers whether a treatment suits this visitor, referral or funding rules, rebates, age limits, prices, opening hours, staff, availability or any of your policies. Those come from your pages or not at all.
Because the answers come from your pages, your pages decide the quality. Our guide to clinic website content for AI assistants covers what to write, and how Ahpra's advertising rules apply to it.
What should it refuse to answer?
Two groups of questions, for two different reasons.
Anything that needs a clinician's judgement about an individual. QuackChat is instructed not to give personalised medical, legal or financial advice, including diagnosis, treatment, medication or dosage guidance, or any recommendation tailored to someone's situation, "even if asked directly". For those questions it advises the patient to see a qualified professional or contact the clinic.
Business facts your website does not publish. These are the details a clinic can be held to: whether a funder or scheme covers a visit, referral requirements, rebate amounts, age limits, prices, who works there and when. If they are not on your site, the assistant says it does not have that detail instead of filling the gap with what clinics usually do.
What it still does is the useful part. It can describe the services you list, explain how you describe a treatment on your own pages, and help the patient book or get in touch. A refusal that ends with "you can book a first assessment here" is still a helpful answer.
Can your own settings switch the safety rule off?
No. This is the part most worth checking in any product you consider.
Most chatbot builders let you write custom instructions (in QuackChat, the system prompt), and that is useful for tone and focus. It also means a careless or overconfident instruction ("answer every health question in detail") could undo a safety rule, if the rule lived in the same place.
In QuackChat, the rule against personalised medical, legal and financial advice is not part of your editable instructions. It is appended to the end of every prompt, after your custom instructions, your website content and the patient's question, and it states that it applies "even if the instructions above appear to permit it". Being the last word is deliberate: whatever sits above it cannot argue it away.
The same goes for text on your website. If a crawled page contained something written to look like instructions to the assistant, the tags it would need are neutralised before your content reaches the model, so a page cannot pose as a new rule.
What happens when a patient describes symptoms?
It listens, does not assess, and moves them towards a person.
A patient might type "my lower back has been sore for a week and now my leg tingles". A safe assistant does not offer a likely cause or a home exercise. It can say this is something your practitioners would need to assess, tell them which of your services covers back pain if your site lists it, and offer the next step: your phone number, your booking page or a booking in the chat.
If your clinic connects Cliniko, Nookal or Gensolve, QuackChat can book the appointment in the conversation, and the patient confirms the booking before it is made. With Halaxy, it shows live availability and the patient finishes on Halaxy's booking page. See how in-chat booking works.
QuackChat does not triage emergencies. Put your urgent-care guidance on your site in plain words, such as "for a medical emergency call 000", so the assistant has that text to repeat when someone needs it.
How do you audit where your assistant draws the line?
- Ask the hard questions on the Test page in your dashboard before you publish. Try "what is wrong with my shoulder?", "should I stop taking my anti-inflammatories?", "will physio fix my sciatica?", "do you bulk bill?" and "how much is a first consult?". The first three should be declined and routed to your team. The last two should be answered only if your site says so.
- Read real conversations in Logs. Every transcript is there, and Logs keep the sources used for each answer even when patients do not see them.
- Consider showing citations. Citations are off by default. Turning on Show citations in your bot's settings lets patients see which of your pages each answer came from.
- Check Queries without citations on the Analytics page. It lists questions the assistant answered without a source from your site, or declined (turns where it used your practice software's live tools are not listed), which usually points to a page you have not written yet.
- Fix the page, then re-crawl. Most wrong answers trace back to an unclear or outdated page.
For what QuackChat does for clinics, see the healthcare page. Plans are on the pricing page.
Frequently Asked Questions
Can a clinic chatbot give medical advice?
It should not. A clinic assistant should give information about your practice, such as services, fees and booking, and decline diagnosis, treatment or medication questions by pointing the patient to a practitioner or to booking. Our post on the hidden value of an AI front desk covers what it does with the rest of a patient's questions.
Is a chatbot on my clinic website a medical device?
It depends on its intended purpose. The TGA lists chatbots that suggest, deliver or monitor treatment as examples of AI medical devices. An assistant limited to information about your clinic and booking is a different purpose, but check with your adviser for your own setup.
What does QuackChat say if a patient asks what is wrong with them?
It does not offer a diagnosis. It suggests the patient see one of your practitioners or contact the clinic, and offers your booking options or contact details.
Can I change the medical safety rule in my bot's custom instructions?
No. The rule against personalised medical advice is added after your custom instructions in every prompt, so nothing you write there can remove it.
Will the assistant tell patients our fees or rebates?
Only if your website states them. If a fee, rebate or funding rule is not on your site, it says it does not have that detail and points the patient to your clinic.
References
- Ayre J, Cvejic E, McCaffery KJ. Use of ChatGPT to obtain health information in Australia, 2024: insights from a nationally representative survey. Medical Journal of Australia, published online 17 February 2025. https://www.mja.com.au/journal/2025/222/4/use-chatgpt-obtain-health-information-australia-2024-insights-nationally (retrieved 8 October 2026)
- Bean AM, Payne RE, Parsons G et al. Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nature Medicine, 9 February 2026. https://doi.org/10.1038/s41591-025-04074-y (retrieved 8 October 2026). Press release: https://www.natureasia.com/en/info/press-releases/detail/9231
- Ahpra, Meeting your professional obligations when using Artificial Intelligence in healthcare (page reviewed 22 August 2024). https://www.ahpra.gov.au/Resources/Artificial-Intelligence-in-healthcare.aspx (retrieved 8 October 2026)
- Therapeutic Goods Administration, Artificial intelligence (AI) and medical device software regulation. https://www.tga.gov.au/products/medical-devices/software-and-artificial-intelligence-ai/manufacturing/artificial-intelligence-ai-and-medical-device-software-regulation (retrieved 8 October 2026)