A health claim can travel through a family chat or social feed faster than anyone checks it. A confident tone, screenshot or citation does not establish that the claim is accurate. The useful question is: what exactly is being claimed, and does a suitable current source support that wording for that population and context?
This method is for public, general health claims. It keeps private messages and personal health details out of AI tools, and it keeps the final judgment with the reader and an appropriate professional.
Do not use this exercise for your symptoms, medicines, treatment or an urgent concern. Contact a pharmacist or clinician through an official local route. For immediate danger or signs of an emergency, contact local emergency services. A source trace cannot decide personal care.
Why chatbot answers are not source verification
A 2026 BMJ Open audit tested five chatbots across 250 adversarial prompts in five misinformation-prone health areas. In that specific audit, independent experts rated 49.6% of responses problematic, and the median completeness score for the chatbots’ cited references was 40% (Tiller, Marcon, Zenone et al., 2026). These figures describe the five tested products, prompts, date and scoring method. They are not a universal error rate for all chatbot health responses.
The result does establish a practical boundary: a fluent chatbot answer or citation list is not verification. Open and assess the source yourself.
Before you use a tool
Do not paste a group-chat message, screenshot, sender name, profile, personal story or health detail into a general-purpose AI service. Retype only the public, checkable claim in neutral words. Removing names does not make a private message anonymous.
You can complete the whole worksheet manually. If you use AI at all, use only the bounded query-writing prompt below. Do not ask it to find sources, summarize evidence or decide whether the claim is true. Turning model training off does not separately control retention, history, human review, sharing or connected apps.
Step 1: Write one checkable public claim
Strip away urgency, endorsements and calls to action. Keep quantities, population, intervention and outcome if they are part of the claim.
Example:
Public claim to check: "Vitamin X reduces the duration of the common cold
in adults."
If the claim cannot be stated without revealing someone’s private health information, do not put it into an AI tool. Work manually or ask an appropriate professional.
Step 2: Build neutral search queries
Start with your national health authority or medicines regulator, a relevant clinical guideline, a systematic review, or the original research. If you need help phrasing a search, this is the only AI prompt in the method:
This is a public, non-personal health claim: "[retyped claim]".
Write three neutral search queries: one for an official public-health or
medicines authority, one for a current clinical guideline or systematic
review, and one for the original research. Do not name sources, browse,
summarize evidence, judge the claim, add facts or take actions.
The instruction not to browse or act is not a security control. Tool access and data handling must enforce the same boundary.
Step 3: Search and open the sources yourself
Use a normal search engine, an authority’s own site or a research index. Do not count a citation until you open it. Record the author or publisher, purpose, publication or update date, population, intervention or exposure, outcome, qualifications and evidence design.
MedlinePlus recommends checking who runs and funds a source, its purpose, evidence, review process, date and privacy practices, then comparing the information with other reliable sources. Its evaluation checklist turns those questions into a short review.
If you cannot locate a source, mark the claim unsupported in this review. A missing or fabricated citation does not by itself prove that every possible version of the underlying claim is false.
Step 4: Check exact support and applicability
Compare the claim word by word with what the source actually says. Ask:
- Is the population the same?
- Are the dose, duration, outcome and comparison the same?
- Is this one study, a guideline or a synthesis of several studies?
- Is the source current, and does newer guidance change it?
- Are limitations, funding or conflicts disclosed?
- Do suitable sources disagree, and on what point?
Do this reading yourself. A chatbot paraphrase could omit the very limitation you need to see.
Step 5: Record a bounded decision
Choose one:
- Share with the source and the context required to keep the wording accurate.
- Do not share because the checked source does not support the wording.
- Do not share while support or applicability remains uncertain.
The health claim source-trace worksheet keeps the source chain and decision together. A source check does not authorize changing a medicine, treatment, diet or care plan.
Common pitfalls
- Pasting the original private message instead of retyping a public claim.
- Asking AI to find sources, then treating its citations as checked.
- Asking AI to summarize a source instead of reading the source.
- Generalizing the BMJ Open audit beyond the products and prompts tested.
- Treating one supporting study as consensus.
- Sharing “just in case” when applicability remains uncertain.
Try it today
Take one public health claim, write it neutrally, create search queries, open suitable sources yourself and record the decision. If the claim could affect someone’s care, take it to an appropriately qualified professional rather than relying on the worksheet.
This is health-information literacy, not medical advice or a truth-certification service.
