How do I stop AI from making up answers?
Paste one instruction on the line above your question, before the model answers: tell it to say 'I don't know' when it isn't certain and to flag anything it can't verify instead of guessing. This pre-empts fabricated answers and citations rather than catching them afterward, and it marks the specific claims the model is unsure about.
How to do it
- Start a new line above your actual questionBefore typing what you want to ask, leave the top line free for the instruction. It has to sit above the question, not below it.
- Paste the instruction verbatimCopy it exactly: If you're not certain, say 'I don't know' and flag anything you can't verify — don't guess.
- Write your real question below itAdd your actual question on the next line and send both together as one message.
- Read the flagged parts of the answerTreat anything the model marks as unverified or 'I don't know' as a gap to check yourself, not as a finished fact.
Questions that come up
Does this work on any AI model?
It works on most current chat assistants because it steers behavior through the prompt, not a hidden setting. Results vary by model, and it reduces confident fabrication rather than eliminating every error, so it helps most on obscure or high-stakes questions.
Will it refuse to answer things it actually knows?
Usually not. On well-established facts the model still answers normally; it mainly gets more cautious on shaky or unverifiable claims. If it hedges on something you know is common knowledge, ask it to answer and explain its confidence.
Does it guarantee no hallucinations?
No. It lowers the odds of a confident made-up answer and surfaces what the model is unsure about, but it can still be wrong. For anything that matters, verify the flagged claims against a real source before relying on them.
Full transcript
One line above your question gets your AI to admit uncertainty. Paste this on top: if you're not certain, say I don't know, flag what you can't verify. It steers the model before it answers, and marks claims it's unsure of. Same question: plain, it fakes a citation. With the line, it says I can't verify that. Save that line, paste it above your next question.