A thread about intermittent fasting was tearing through my health-conscious group chat. Nine specific claims, each citing 'studies' it never named, each delivered with the confident specificity that makes nonsense feel like knowledge: 'a 6-hour eating window boosts autophagy by 300%,' 'fasting permanently raises your metabolism,' that genre. Four different friends had forwarded it to me, presumably because I'm the one who 'reads things' and therefore must rule on whether they should buy the meal-timer app at the end. I didn't know what to do with it, and 'just trust the vibe' was clearly not going to cut it when there was a $40 subscription riding on my verdict.
The problem with a thread like this is that it's engineered to bypass your skepticism on purpose. The numbers are oddly precise — 300%, not 'a lot,' because a specific number sounds researched. The tone is calm-expert, never salesy. Each individual claim is plausible enough that stopping to doubt it feels paranoid. By claim seven you've been gently lulled into accepting the framing, and the framing is the actual trick. I needed something that would go claim by claim, in cold blood, and refuse to be charmed by the production values.
Sorting confidence from accuracy
I pasted all nine claims into Grok 4.3 with the Triage a List of Claims by What You Can Actually Verify prompt, context set to 'viral social-media health content I'm deciding whether to share.' I like Grok 4.3 for this because it's fast and current and doesn't hedge itself into uselessness — it gives you a punchy verdict and moves on, which is exactly the right energy for triage. I don't want a six-paragraph meditation on epistemology per claim; I want a verdict, a reason, and a pointer to where the real answer lives. The output sorted the nine: two SUPPORTED, two CONTESTED, three CONTRADICTED, and two NEEDS CONTEXT — true only under conditions the thread had conveniently skipped over.
The most valuable verdicts were the NEEDS CONTEXT ones, because those are the claims that win arguments by omission. 'Fasting lowers insulin' is technically true and aggressively misleading without the missing half: 'because you haven't eaten, like literally anyone who skips lunch, and it goes back up when you do.' The nuance note named the missing condition in a single line. That's the most contagious category of wrong, because it's not false enough to debunk in a snappy reply and not true enough to actually act on — it just sits there, technically defensible, quietly steering people.
- The verbatim disclaimer up top kept me honest: this is a triage, not a verdict
- Nuance notes caught the technically-true-but-misleading claims — the most contagious kind by far
- Low-confidence ratings were genuinely admitted, not buried, which told me precisely where I'd need to dig myself
- It pointed me to source types — peer-reviewed meta-analyses via PubMed, medical-society guidelines — instead of pretending to be the final word
- The CONTRADICTED claim about metabolism came with the general direction of the actual evidence, so I knew the correction wasn't just 'nuh-uh'
What I actually did with it
I didn't share the thread, and I didn't buy the meal-timer app. I did send the three most persuasive friends a short note explaining which claims were bent and why, with the honest caveat that I'd triaged rather than adjudicated — and for the two claims that genuinely mattered to their health, I pointed them at where to read further rather than declaring the answer myself. Two were grateful. One informed me that I was 'no fun at parties,' which is fair and probably true, but I'd genuinely rather be the no-fun friend than the one who launders a 300% autophagy statistic into the group chat and gets quoted on it for a year.
The limit it forces you to remember
The honest limit, which the prompt forces you to print verbatim at the very top so you can't skip it, is that an AI fact-check is a triage tool, not a tribunal. It tells you which claims are worth verifying and roughly which direction the wind is blowing — it does not, and should not, replace reading the actual meta-analysis for anything you're going to repeat in public with your name attached. I treat it exactly like a metal detector at the beach: it beeps to tell you where to dig, it does not hand you the gold, and anyone who treats the beep as the treasure is going to end up with a pocket full of bottle caps.
Next time a suspiciously confident thread lands in your feed with three friends already converted, grab the Triage a List of Claims by What You Can Actually Verify prompt on Prompt Dock, paste the claims in, and let Grok 4.3 tell you which ones are real before you become the reason your group chat believes them.