Erin is in her early 40s. She works part time in a demanding job and has two kids in primary school. She feels constantly exhausted, she's often angry and feels like she can't think like she used to. Her friend tells her it's perimenopause.
Authors
- Kirsten McCaffery
NHMRC Principal Research Fellow, Sydney School of Health, University of Sydney
- Julie Ayre
Post Doctoral Research Fellow, Sydney Health Literacy Lab, University of Sydney
- Liliana Laranjo
Senior Lecturer in Digital Health and Community and Primary Health Care Practice, University of Sydney
She books an appointment to talk to her GP. She doesn't know anything about perimenopause so asks ChatGPT what it is and what treatments are available.
Erin is certainly not alone. Between 30% and 55% of adults report they have used ChatGPT, or other large language models or chatbots for their health. And this use is increasing .
Even if you're not using an AI chatbot, a simple Google search will often produce an AI summary.
So how does this impact patient care? And how can you make the process safer and more effective?
Why do people use AI chatbots for health?
A recent survey in the United Kingdom found while some people asked chatbots health questions because they were curious, more than one in three used them to decide whether they needed to see a GP.
Around half found chatbots more convenient than regular services, with questions frequently occurring at night or in the morning , outside the hours of traditional health services .
Of those who used a chatbot, more than half believed it had been good for their physical and mental health.
A 2026 analysis of over 500,000 health conversations with Microsoft co-pilot found one in five related to personal symptoms and treatments. Many queries related to someone else, such as a child or family member.
These findings echo results from our own nationally representative Australian survey . More than one-third of those who used ChatGPT for their health did so to find out what symptoms meant. Half used it to learn about a specific health condition.
Health information can be very complex and is often poorly communicated . Consultations are short and patients frequently report not having enough time or confidence to ask clinicians all the questions they want.
By comparison, chatbot advice is readily available, personalised, and can be prompted for greater clarity and simplicity without constraint and fear of judgement.
What are some of the risks?
There are important reasons for alarm if chatbots are used when people need urgent medical attention and the answers they give are unsafe.
In the UK survey described above, one in five respondents decided against seeking medical advice from a professional after consulting a chatbot.
There is currently little real-world data on the safety of people using AI chatbots for health decisions, other than a few individual case studies of harm.
This week the media has also reported on the death of 31-year-old Bowen Wang , who died in February after ChatGPT gave her the wrong advice about cutting back her epilepsy medication.
However, simulation studies examining the accuracy of chatbot health advice show cause for concern .
One study investigated AI responses to 222 patient questions that were relevant to primary care. Between 5% and 13% of the responses were unsafe with potential to lead to "serious patient harm", depending on the AI chatbot used.
Another study found that although chatbots were accurate 94.9% of the time using hypothetical clinical scenarios (with prompts by the research team), in searches operated by study participants (using the same clinical scenarios supported with the same chatbots) accuracy dropped to 34.5%.
It also found using an AI chatbot may be less effective than searching the web when finding out about an underlying condition and was similar to a web search for deciding whether to seek professional care, with only 43% accuracy.
One of the key issues is people often do not provide enough detail in their prompts to obtain accurate results. However, effective prompting is difficult to do (if untrained) and importantly, requires a careful trade-off between specificity and privacy.
Nor do consumers cross-check information, despite sources sometimes being incorrect or irrelevant. Only 1% of users click on the links provided by AI Google summaries.
Patients don't always disclose their use
In an ideal world, patients and clinicians can collaboratively discuss the use of chatbot advice to help support shared decision-making .
But patients vary in their decision to disclose their chatbot use to their doctors. Many don't - between 83% of patients in a Polish survey to 32% in the US.
Reluctance to disclose use was due to fear of a negative reaction or concern the clinician might feel undermined.
But recent US data suggests generally positive views from clinicians about patient AI use under certain conditions.
Of the 1,692 clinicians surveyed, 70% considered use of AI chatbots by patients for their general health information or decision-making was positive or had no impact to their practice.
But around 50% said they would rarely or never want their patient to use AI for interpretation of medical test results/reports.
How to get better outcomes
Disclosing AI chatbot use enables clinicians to correct any inaccuracies and use it in shared decision-making .
But clinicians say they're currently poorly trained and lack time (and reimbursement) to support their patients' AI use.
While some resources are available , patients need much more help in using chatbots, and clinicians need more guidance on how and where to direct them for support.
The new SMART guide, designed by an international team of clinicians, researchers and AI health experts, spells out five key steps to help people use chatbots more safely:
- Slow down and think before asking health questions to a chatbot
- Make your questions clear
- Ask yourself "what am I sharing?"
- Remember to check the facts
- Talk to a real person when you need to.
People also need to know they can access human-led, 24/7 triage services for free, such as 1800MEDICARE in Australia (1800 022 222) and NHS 111 in the UK. These services are essential to ensuring people get the right advice and care when they need it.
Kirsten McCaffery receives funding from the National Health and Medical Research Council (APP2016719). She is a co-director of Health Literacy Solutions Pty Ltd, which holds a sub-licence for the Health Literacy Editor, a research tool developed by researchers at the Sydney Health Literacy Lab and owned by the University of Sydney. The sub-licensing arrangement enables the broader dissemination and public use of the tool. KM receives no personal income or financial benefit from either Health Literacy Solutions Pty Ltd or the Health Literacy Editor. KM is past president of the International Society for Shared Decision Making. She received no income or financial benefit in this role.
Julie Ayre receives funding from the National Health and Medical Research Council, and Diabetes Australia.
Liliana Laranjo receives funding from the National Health and Medical Research Council (APP2017642) and Sydney Horizon Fellowship (The University of Sydney).
