This is not a future problem. It is a current clinical reality that affects consultation dynamics, patient safety, and the clinician's role in verifying and correcting health information.
King's College London reported that 15% of the UK public had used AI chatbots for health advice instead of contacting a GP or other NHS service, and 10% had used AI for mental-health therapy or wellbeing support instead of seeing a trained professional. Twenty per cent of people who sought AI health advice said it did not encourage them to seek professional opinion. Twenty-one per cent decided against seeking professional healthcare advice because of something an AI chatbot said.
In the US, KFF found that 32% of adults had used AI chatbots in the past year for health information — 29% for physical health and 16% for mental health. Forty-two per cent of those who asked about physical health and 58% of those who asked about mental health did not follow up with a doctor or other professional afterwards. Quick and immediate advice was the most common major reason for AI use. Many users consulted AI before deciding whether to see a provider — using AI as a pre-triage filter rather than a post-consultation reference. Younger adults and lower-income users were more likely to cite affordability or access barriers as reasons for using AI instead of professional care.
Why This Is Different from Normal Internet Searching
"Dr Google" returned links to websites. The patient read, compared, and formed their own interpretation. The information was visibly external — a webpage, a forum post, a medical reference site. The patient understood they were reading someone else's content.
AI chatbot health advice is different in several clinically important ways. It is conversational — it feels like talking to someone, not reading a webpage, creating a false sense of personalised interaction. It is confident — the response sounds authoritative even when uncertain, with no visible hedging or source qualification. It is tailored — it responds to the user's specific wording, creating the illusion of personalised medical advice when the response is actually a statistical pattern-match against training data. It is action-oriented — AI may suggest specific actions (starting supplements, stopping medications, waiting rather than seeking care, performing self-examinations) rather than simply providing information. It is detached from local pathways — AI does not know the user's GP, NHS service options, local referral routes, or regional service availability. It is difficult to audit — unless the patient shows the clinician the AI output, the clinician has no way to know what advice was given, how it was framed, or what it omitted.
What Clinicians Should Ask
Incorporating AI use into clinical history is practical and non-confrontational. These questions turn AI use into clinical information rather than a confrontation about self-diagnosis.
"Have you looked this up online or asked an AI tool about it?"
"What did it tell you?"
"Did it make you more worried, less worried, or change whether you came today?"
"Did it advise you to start, stop, or change any treatment?"
"Did it suggest you didn't need to see a doctor?"
"Would you be happy to show me what it said, so we can check it together against your specific situation?"
The last question is particularly valuable — if the patient shares the AI output, the clinician can see exactly what was said, identify what was correct, what was incomplete, and what was wrong, and use it as a teaching moment that strengthens rather than damages the therapeutic relationship.
Red-Flag Scenarios Where AI Advice Is Particularly Risky
AI-generated reassurance may delay presentation for time-critical or safety-critical conditions. Clinicians should specifically ask whether AI advice influenced the timing or nature of help-seeking in any of these scenarios:
Chest pain — AI may reassure "probably muscular" without adequate assessment of cardiac risk factors. Breathlessness — AI may attribute to anxiety without considering PE, heart failure, or pneumothorax. Focal neurological symptoms — AI may suggest migraine aura without considering stroke or space-occupying lesion. Pregnancy — AI may provide medication advice that contradicts UK SmPC pregnancy safety data. Infants and young children — AI may underestimate the significance of fever, poor feeding, or behavioural change in young children. Immunosuppressed patients — AI may not recognise that infections in immunocompromised patients require different assessment thresholds. Severe or sudden-onset headache — AI may normalise thunderclap headache without considering subarachnoid haemorrhage. Suicidal ideation or self-harm — AI mental health advice may be inadequate for acute risk assessment and may not direct the user to crisis services. Medication changes — patients acting on AI advice to stop anticoagulants, reduce insulin, or discontinue psychiatric medication without professional guidance. Cancer red-flag symptoms — AI may normalise weight loss, rectal bleeding, or persistent cough without prompting urgent assessment. Safeguarding concerns — AI cannot assess non-verbal cues, home environment, or relational dynamics that indicate abuse or neglect.
How to Correct AI Advice Without Losing the Patient
The effective approach validates the patient's initiative while establishing clinical authority:
"It makes sense that you checked — a lot of people do now. Some of what it told you is reasonable, but there are a few things AI cannot assess properly in your situation: your examination findings, your specific risk factors, your other medications, and the local guidance we follow here. Let me explain what I'd recommend based on what I can see and what we know about you specifically."
This avoids sounding dismissive ("AI is unreliable, don't use it") while making clear that the clinician's assessment is based on information the AI does not have — examination, full clinical history, patient-specific context, and jurisdiction-specific guidelines.
Documentation
When a patient reports using AI for health advice, consider documenting: that the patient used an AI chatbot for health advice (and which one, if known), what advice they report receiving (in their words), whether it delayed or changed their help-seeking behaviour, whether medication advice was given and whether the patient acted on it, the clinician's independent assessment (based on history, examination, and clinical judgement), safety-netting provided (specific, time-bound, presentation-appropriate), red flags explained to the patient, and the follow-up plan.
This documentation protects the clinician medico-legally, creates a record of how AI-mediated health information is affecting patient behaviour, and may be important for clinical governance monitoring at practice and system level.
Where iatroX Fits
Patients will increasingly arrive with AI-generated explanations. Clinicians need fast, source-grounded tools to verify or correct those explanations against recognised guidance, drug information, and clinical context. iatroX provides cited UK clinical answers that clinicians can use to check, confirm, or correct AI-generated health advice — with visible provenance, source grounding, and the professional accountability framework that consumer AI lacks.
Use iatroX to verify clinical information when patients arrive with AI-generated health advice →