ShareYourGroupShareYourGroupAI Prompts

Guides

How to check an AI answer before you use it

An AI assistant writes fluent, confident text whether or not the content is correct. That is the single most important thing to know about using one. This guide explains what goes wrong and gives a simple routine for catching it.

Why confident mistakes happen

An assistant produces text that is likely to follow from your prompt, based on patterns in the enormous amount of writing it was trained on. Often the likely text is also true. Sometimes it is not, and the assistant has no reliable inner sense of the difference. It does not hesitate when it is on weak ground. A wrong answer arrives in the same assured tone as a right one.

The kinds of mistake to expect

What to check, by how much it matters

Not everything needs the same care. Sort the answer by what happens if it is wrong.

What you are using it forHow to treat the answer
Wording, tone, structure, ideasUse freely. Read it for sense.
A summary of text you pastedCompare with the original for missing or altered points.
Facts, names, dates, numbersCheck each against a reliable source before using.
Rules, fees, deadlines, proceduresConfirm on the official website or with the office.
Health, law, money, safetyTreat as questions to ask a qualified person. Do not act on it alone.

A checking routine

  1. Read it all, slowly. Most errors are visible to a careful reader who knows the subject. That reader is you.
  2. Mark every checkable claim. Names, dates, amounts, figures, quotations, references.
  3. Check each against something that is not the assistant. Your own records, the original document, an official website, a textbook.
  4. Recalculate the numbers. Use a calculator or a spreadsheet.
  5. Look for what is missing. A summary can be accurate in every sentence and still leave out the one thing that matters.
  6. Ask yourself who will rely on it. The more people, the more checking.

Asking the assistant to help you check

Some prompts make checking easier.

List every factual claim in your last answer that I should verify, such as names, dates, numbers and rules. For each, tell me how sure you are and where a person could confirm it. Do not give web addresses unless you are certain they exist.

Replace the parts in [CAPITALS] first, or after pasting.

Which parts of your answer came directly from the text I pasted, and which parts did you add from your own general knowledge? Mark the added parts.

Replace the parts in [CAPITALS] first, or after pasting.

These help, but they are not proof. An assistant asked "are you sure?" may simply repeat its mistake politely, or change a correct answer because you sounded doubtful. Its own statement of confidence is not evidence.

Do not ask it to confirm itself

Opening a second conversation and asking the same question is a slightly better test, since a made-up detail often changes between attempts. But two matching answers can both be wrong. The only real check is a source outside the assistant.

Where to look instead

The right attitude

Think of the assistant as a fast, well-read helper who never says "I do not know". Such a helper is valuable for drafts, ideas and explanations, and dangerous as a final authority. Use it to do the first eighty per cent quickly, and spend the time you saved on checking the part that matters.

This is an independent site. It is not connected with any company that makes AI assistants. The prompts are plain text and can be used in any AI chat assistant. An AI assistant can give wrong or made-up answers: check anything important before you rely on it, and do not paste private details into a prompt.