AI Says It Knows What’s Wrong. Should You Trust It?

AI says it knows what is wrong. Treat that as a possible explanation, not a diagnosis. A useful medical answer should help you describe symptoms and prepare questions; it cannot replace examination, medical history or a clinician's responsibility for a decision.

The World Health Organization's guidance on large multimodal models in health describes possible uses in care and research alongside risks such as inaccurate or incomplete statements and biased data. The practical challenge for a patient is to recognise which parts of an answer are general information and which would change what they do next.

Start with urgency, before debating the bot's explanation

If symptoms might require urgent help, use the local health service's emergency route. Do not spend time persuading a chatbot to agree that they are serious. Warning signs include difficulty breathing, chest pain, bleeding that will not stop, fainting, a sudden change in mental state or sudden weakness. This is not a complete symptom list or a way to diagnose yourself; MedlinePlus lists emergency warning signs and advises prompt help when a condition may be life-threatening.

In England, the NHS says to call 999 for a life-threatening emergency and use NHS 111 when urgent medical help is needed but it is not an emergency. In the United States, call 911 for an emergency. Throughout the EU, 112 reaches emergency services; national urgent-care routes vary. Elsewhere, use the local emergency number. If you are unsure whether a symptom is dangerous, seek professional triage promptly instead of waiting for an AI answer.

Check the official advice for your location: NHS emergency guidance, the US National 911 Program and the European Commission's 112 guidance.

An apparently calm response does not establish that a symptom is harmless. A bot may not know your age, medication, pregnancy status, history or what a clinician would find on examination. The absence of those details can change the appropriate next step. If a person is getting worse, the current situation matters more than the chatbot's earlier answer.

Separate a possibility from a conclusion

Ask what evidence supports the suggested condition and what alternative explanations could fit. A symptom such as tiredness has many potential causes. An answer that names one cause without saying what information is missing has not resolved the problem. Asking the bot for a more confident version of the same explanation does not supply the missing evidence.

Consider a hypothetical example: someone describes a persistent headache and asks for a cause. The chatbot lists common triggers and offers a neat label. A better use is to make a note of onset, duration, associated symptoms, medication and relevant history to share with a clinician. The example does not establish what caused any real person's headache or how urgently they need care.

Check any proposed action against an authoritative source

Be especially cautious if an answer tells you to start or stop medication, delay an appointment, dismiss a warning sign or use an unproven treatment. Look for a current source from a recognised health authority that applies to your country. Then take questions about your situation to a qualified professional. Even an accurate population-level statement may be unsuitable for one person.

To check a chatbot's claim, isolate the sentence that would change your action. Open the health authority's own page rather than a link summary, check when it was last reviewed, and read its warnings and eligibility conditions. Compare the claim with another reliable clinical source if the decision matters. If the sources disagree, your symptoms are changing, or the answer would alter treatment, call a local health advice line or contact a clinician; do not ask the chatbot to break the tie.

WHO's 2026 report on responsible AI in health identifies governance, accountability and biased data as barriers to safe deployment. This is a reminder that model quality is not merely a matter of a convincing interface. A tool integrated into a clinical service may have a defined purpose, oversight and evaluation; a general chatbot may not.

Protect your health information

Medical details are sensitive. Before pasting records, photographs or identifying information into an AI product, check who operates it, how the data may be used, and whether you can delete the conversation. If a general explanation is all you need, describe the issue without names, addresses, account numbers or full records. A privacy check is separate from checking whether the health answer is correct.

Do not upload a family member's records merely because you have access to them. They may not expect an outside service to process those details. For your own notes, keep the description to what is needed for the question, and check the product's current privacy terms. If a clinician or health service has supplied a secure patient portal, follow its instructions for sharing records there. Removing a name from a document may not remove other identifying clues. When you are unsure, ask the service how to send the information safely.

Privacy is one reason to pause; accuracy and urgency are others. A private tool can still produce a wrong answer, and a well-sourced general answer can still miss your circumstances.

Bring a better question to the appointment

Use AI to organise what you want to tell a clinician: what happened, when it began, what has changed, which treatments you have tried, and what you most need to understand. Bring the notes as your account, not as a diagnosis to defend. Ask the professional what would change their assessment and what signs mean you should seek help again.

The safest payoff from a chatbot is often a clearer conversation with a person who can assess you. If the answer would alter treatment or delay care, verify it through the appropriate health service first.