Somebody in the family group chat shares a claim about a supplement, a food, or a treatment, with a screenshot or a confident-sounding paragraph attached. It sounds plausible. It might even be partly true. The question that decides whether to forward it, ignore it, or push back on it is not “does this sound right” - it is “where did this actually come from, and does that source still say what the forward claims it says.”
This is a short method for tracing a health claim before you act on it, using AI to find and organize sources - not to decide whether the claim is true. That distinction matters more here than almost anywhere else AI gets used, because chatbots are measurably unreliable at exactly this task.
Why this needs a method, not a gut check
A 2026 audit published in BMJ Open tested five widely used chatbots - Gemini, ChatGPT, Grok, DeepSeek, and Meta AI - against adversarial health questions in misinformation-prone areas: cancer, vaccines, stem cells, nutrition, and athletic performance. Nearly half of all responses, 49.6%, were rated problematic by independent experts, and citations were frequently incomplete or fabricated outright - the median completeness score for a chatbot’s own cited references was only 40% (Tiller, Marcon, Zenone et al., BMJ Open, 2026). The chatbots delivered nearly all of this with confident, unhedged language: across 250 questions, they refused to answer only twice.
That combination - confident tone, plausible phrasing, and a real chance the cited source does not say what the answer claims - is exactly the profile of a forwarded health claim in a family chat. A model asked “is this true” will often answer as if it can adjudicate the claim outright, when what it is actually doing is generating a plausible-sounding response, sometimes backed by a citation that does not hold up when you check it.
This method is for general claims you encounter in chats, articles, or social media - not for urgent personal questions about your own health or a medication you are taking. Those go to a pharmacist or your clinician, not to a source-trace exercise, however well-sourced the general claim turns out to be.
Step 1: Separate the claim from the packaging
Before checking anything, write down the specific, checkable claim - stripped of the framing that made it feel urgent or exciting.
Here is a health-related message I received: [paste it]. Extract the
single specific, checkable factual claim being made, in one plain
sentence. Ignore the tone, urgency, and any call to action - I only
want the underlying claim.
A message that reads “Doctors don’t want you to know this simple trick!” might reduce to a specific, testable claim like “Vitamin X reduces the duration of the common cold.” That is the sentence you actually need to check.
Step 2: Ask the model to find sources, not verdicts
Ask explicitly for sources and their type - not for a true/false judgment. This keeps the model in the lane where it is actually useful: search and organization, not adjudication.
For this claim: "[claim from step 1]" - find what kind of sources
typically address this topic (health authorities, peer-reviewed
research, professional medical bodies) and suggest specific
organizations or publications I should check directly. Do not tell me
whether the claim is true - I will check the sources myself.
Step 3: Check the source directly - do not trust the summary
Go to the actual source the model named. This step cannot be skipped or delegated, because the BMJ Open findings above show chatbots regularly cite sources that do not fully support the claim, or fabricate details in the citation itself. If the model names a specific study, a health authority page, or a named organization, open that source yourself and read what it actually says.
If you cannot find the specific source the model named - the study does not exist, the page does not say what was claimed, or the organization never published it - treat that as a strong signal the whole claim is unreliable, not as a gap to paper over with another AI query.
Step 4: Check whether the source still applies
A claim can be sourced accurately and still be outdated, population-specific, or lifted out of context. A study on a specific age group, dose, or condition does not automatically generalize to everyone who reads the forward.
Here is what the source actually says: [paste the relevant section].
Help me identify: what population or context this applies to, how
recent it is, and any qualifications the source itself states. Do not
add a conclusion about whether the original claim is fair - just
surface what the source itself says and doesn't say.
Step 5: Decide, and label the decision
With the claim, the source, and its applicability in hand, make one of three calls: share it as-is because it holds up, share it with the missing context attached, or do not share it because the source does not support the claim. Write the decision down along with why - a documented “do not share, source does not exist” is worth more to the next person in the chat than silence.
The full health claim source-trace worksheet walks through all five steps with space to record the original post, the claim, the source chain, and your final decision.
Common pitfalls
- Asking the model “is this true” instead of “find the sources.” The first invites a confident guess; the second keeps the model in the part of the task it is actually good at.
- Trusting a citation without opening it. A citation that reads authoritatively can still be incomplete or fabricated - check it yourself before you rely on it.
- Forwarding “just in case.” A claim you have not traced is not neutral to share; it costs the next reader the same verification work, multiplied by everyone they forward it to.
- Treating one aligned source as consensus. A single supportive result is not the same as checking whether major health authorities agree - look for where the weight of evidence actually sits.
- Skipping the applicability check. A real study about a specific group or dose does not automatically mean the general claim in the forward is fair.
When it is not a sharing decision at all
If the claim is about your own symptoms, a medication you take, or something urgent, stop the source-trace exercise and take the question to a pharmacist or your clinician directly - see AI is not a doctor for why a chatbot should never be the final word on a personal medical question, sourced or not. This method is for claims you are deciding whether to believe or forward, not for decisions about your own care.
Try it today
The next time a health claim lands in your inbox or a group chat, run it through the five steps before you react to it: extract the specific claim, ask for sources rather than a verdict, check the source directly, confirm it still applies, and record your decision. Use the health claim source-trace worksheet to make the habit stick past the first time.
