My grandson showed me how to paste a research paper into one of these AI chat assistants, and I have been checking their homework ever since. Here is what they get wrong. They summarise the abstract and call it the paper. They round numbers. They answer confidently about tables they never read. When there's a gap, they fill it with something that sounds right.

These twelve prompts make an AI slow down. Each one asks for one thing: what kind of study it was, who was in it, what they were compared with, the two numbers behind a risk, who paid, the limits the authors admit. Each one makes the AI quote the paper and say where the quote is. Each one makes it say "not in the paper" instead of guessing, keep its own opinion apart, and stay out of medical advice.

A prompt makes a good answer more likely. Only checking makes it safe to use. So half of this pack is a 10-page guide to checking an AI's answer against the paper in about two minutes: find the sentence, match the number, test the "not in the paper", spot-check a table, and watch for rounding and relative-only numbers.

The guide works through one real, open-access paper from start to finish: a randomised trial of parachutes from the BMJ's Christmas issue. Its abstract says parachutes made no difference. Two sentences later it tells you why. You'll see the prompt, an answer with a mistake in it, and the check that catches the mistake. Then a short list of what the prompts can't do.

Works with the common chat assistants. Paste the paper, then the prompt. Never paste your own health records; the README explains why.

It won't tell you what a paper means for your body. Nothing that has never met you can.