A nutrition study detonated across my feed last spring: 'Intermittent fasting linked to 91% higher risk of cardiovascular death.' Ninety-one percent. That's the kind of number that ends dinner-party conversations and starts group-chat panics. People I love texted me screenshots and announced, mid-sentence, that they were quitting their eating windows effective immediately. I read the actual methods section — which took a careful, coffee-fueled twenty minutes — and felt that specific itch you get when something is wrong but you can't yet name it precisely enough to say out loud. I am a curious layperson, not a cardiologist or an epidemiologist. My gut said 'this is overcooked', but a gut feeling loses every argument against a percentage. I needed a sharper tool.
Pasting it in
I dropped the methods description into the Critique a Research Method and Identify Its Hidden Confounds prompt, set the field to 'nutritional epidemiology', and stated the claim as 'time-restricted eating increases cardiovascular mortality'. GPT 5.2 led with a one-line verdict — 'Does not support a causal claim' — and then spent the next two screens earning it. Its very first confound was the one the press release somehow never found room to mention: the study classified people as 'fasters' if a single day's dietary recall showed them eating within an 8-hour window. But people eat in a tiny window for a hundred reasons that have nothing to do with a chosen health protocol — being acutely sick, frail, hospitalized, recovering from surgery, or too depressed to eat. Illness shrinks your eating window, and illness raises your death risk. That's textbook reverse causation: the outcome is quietly causing the exposure, and a single-day snapshot can't tell the two apart. The design simply could not separate 'I fast on purpose' from 'I barely ate that day because I was unwell'.
The critique didn't stop at one flag
- Confound rated HIGH and controllable: 'pre-existing illness shrinks eating windows and independently raises mortality' — flagged as able to produce the entire association on its own.
- Construct validity problem: using ONE 24-hour recall to label someone's years-long eating pattern. The thing being measured ('habitual fasting') isn't the thing recorded ('what you ate yesterday').
- Missing information it demanded: were the fasters adjusted for baseline disease, weight loss in the prior year, and smoking? Without that, the number is uninterpretable.
- A feasible fix: enroll healthy people first, confirm intentional fasting prospectively, then follow them — and it noted such cohorts already exist, so this wasn't a fantasy.
What I appreciated most is that the prompt did not tell me the study was garbage, and it didn't let me off the hook with a comforting 'fasting is fine, ignore this'. It told me the measured association might well be real, but the causal story the headlines sold — fasting reaches into your chest and damages your heart — was unsupported by this particular design. That distinction is the entire game in epidemiology, and almost every piece of coverage flattens it into a single scary verb. 'Linked to' became 'causes' somewhere between the journal and my group chat, and nobody noticed the swap.
What I did with it
I sent the critique to the three friends who'd panicked the hardest. Two of them actually went back and read the methods section, which I consider a minor miracle of modern friendship. Nobody overhauled their entire diet on the strength of a single-recall observational study, which is a small public-health victory won entirely in my living room. The third friend, to be fair, still thinks I'm insufferable, and that's a fair price. I now run every headline study through this prompt before I let it touch my actual behavior, because the gap between what a study measured and what a journalist wrote about it is where most bad health decisions are born. When I then want to explain the underlying mechanism to someone — say, what time-restricted eating even does metabolically, regardless of this study — I switch over to my Explain Any Scientific Concept With a Concrete Analogy and Its Limits prompt and let it build the picture from scratch.
Grab the Critique a Research Method and Identify Its Hidden Confounds prompt on Prompt Dock and run it on the next study that hijacks your group chat at 7am. Paste the methods section, name the claim the way the headline framed it, and watch it cleanly separate 'the data the researchers collected' from 'the dramatic sentence a journalist later wrote about the data'. Read the severity ratings and the 'feasible fix' section especially — they teach you, over time, to spot the same patterns yourself, until one day you read a scary headline and the confound just jumps out at you before you've even opened the prompt. It is, without exaggeration, the most useful sixty seconds I spend on science news in any given week, and it has quietly made me a much calmer person to follow on the internet.