Guardrails · Method
Make it show its working
An AI hallucination is a model blending what it knows with what it invents, in one tone.
- 3 min read
- Start Here 3 of 4
This site is the AI Reliability Tracker: we check six AI apps’ answers every month.
An AI doesn’t tell you which parts of its answer it’s sure about. It hands you the facts and the guesses in the same even voice, packed into the same tidy paragraph, and leaves you to work out which is which. Usually you can’t. The guesses wear the same face as the facts. When one of those guesses is wrong, the technical word for it is a hallucination, though “confident guess” describes it better.
So I stop it before it gives me an opinion and make it do one thing first: sort what it just read into two lists. List A is only what the document says. List B is everything it’s adding on top: the assumptions, the filled-in gaps, the bits it’s inferring. Paste this in with whatever you’re checking:
Before you give me any opinion or summary, sort your answer into two lists. List A: only what this document says, in its own words. List B: anything you are adding, assuming, inferring or guessing that the document does not state outright. If you are unsure which list something belongs in, put it in List B.
Why two lists beats one summary
This is the one move I’d keep if I had to drop everything else, and it’s the filter stage of the Prompt Stack in a single prompt. A summary mixes everything together and smooths the joins. Two lists pull the joins apart. The second you can see List B on its own, the guesses stop hiding. They’re sitting in a column with a label on it.
Take a letter from your doctor. You paste it in, ask what it means, and the AI says, calmly, “this is nothing to worry about.” Reassuring. Also not in the letter. Run the two-list prompt and that line lands in List B, where it belongs: the comfort it invented, sitting in the guesses column with a label on it, next to everything else it added to be kind. The letter never said don’t worry. The AI did, because it sounded like the helpful thing to say.
That’s the whole value of the move. List A you can check against the document, line by line. List B is where you slow down: some of it will be fair inference, some will be the AI being agreeable, and now you can tell the two apart instead of swallowing both.
It works on anything you’d paste in: a contract, a job offer, an email from your child’s school, a holiday booking. List B is the part that’s a question wearing the clothes of an answer, one of the recurring ways AI gets it wrong.
Next: where I use all of this → Where I actually use this
Free, with Confidently Wrong every fortnight.
Confidently Wrong, every fortnight.
Which AI told the truth this month, with the receipts, plus an AI story worth retelling in between.
Free. Leave whenever you like.