Polling for King’s College London finds convenience and curiosity outrank NHS waiting times as the reason people turn to a chatbot over a doctor — and one in five of those who have say it talked them out of seeing a professional.
Fifteen percent of UK adults — one in seven — have used an AI chatbot for health advice instead of contacting a GP or other NHS service, according to a Focaldata survey of 2,093 UK adults for King’s Health Partners, Responsible AI UK and the Policy Institute at King’s College London, fielded 24–30 March 2026.1 A further one in ten (10%) have turned to a chatbot for mental health support instead of seeing a trained professional.1
The obvious explanation is an overstretched NHS. It is a real factor: a quarter (25%) of people who have used AI for health advice cite long waits for NHS services as a reason.1 But it is not the leading one. Convenience (46%) and plain curiosity (45%) are cited far more often, and uncertainty about whether a symptom was serious enough to bother a GP (39%) also outranks the waiting list.1 Respondents could pick up to three reasons, so the shares do not sum to 100 — but the gap between the top reason and NHS waits is far wider than sampling error would allow.
Among people who have used AI for health advice, one in five (20%) say the chatbot never encouraged them to seek a professional opinion, and a near-identical share (21%) say they decided against seeking professional care because of something the AI told them.1
That risk lines up with an independent finding on how these tools perform in exactly this situation. A JAMA Network Open study published in April 2026 tested 21 large language models — including consumer chatbots from the major developers — against 29 standardised clinical vignettes.2 Every model’s failure rate exceeded 80% at the differential-diagnosis stage: the early point where symptoms are still incomplete and several conditions remain plausible — the situation of someone deciding whether a symptom is “serious enough” to bother a GP. Accuracy improved sharply once the full clinical picture was available, with failure rates below 40% for a final diagnosis.2 Reasoning-optimised models scored higher on balance but did not close the gap.2
The survey captures self-reported behaviour, not outcomes — it cannot say how many of the AI-instead-of-GP episodes involved something that genuinely did not need a doctor, and how many did. “AI chatbots” is also one broad category here, covering general-purpose tools like ChatGPT, Copilot and Gemini rather than dedicated symptom-checkers built for triage. And the JAMA study tested models with their reasoning settings switched off and used a public medical-reference vignette set that published critiques note could overlap with model training data; it did not compare any model against a doctor working the same cases. So its figures describe model performance in isolation, not a measured AI-versus-clinician gap.