How to Stop AI Making Things Up: A Verification Checklist

You cannot prompt your way to zero errors. Anyone who promises a prompt that "stops AI hallucinations" is overselling. What you can do is make wrong answers rarer, easier to spot and cheaper to catch. That is a realistic goal, and it is enough to use AI safely for most everyday work.

This article explains why assistants state wrong things, shows two prompting habits that reduce the problem, and gives you a checklist for verifying the details that matter: names, numbers, quotes and citations. It is general information, not professional advice.

Why AI assistants state wrong things with confidence

A chat assistant generates text that fits the pattern of what you asked. It is built to produce a fluent, plausible reply, and fluent does not mean true. Unless a search or document tool is attached, it is not looking anything up. It is predicting what a good answer would look like.

That is why a wrong answer often sounds identical to a right one. The tone does not change when the facts do. There is no internal alarm that rings when it is guessing, so you need an external habit that does the job instead.

Where made-up answers show up most

Invented content tends to appear in predictable situations. Knowing them tells you where to aim your checking.

• You did not supply the material. Ask for "the three main findings of the report" without pasting the report, and you will likely get three findings anyway.
• The detail is exact and obscure. Quotes, page numbers, citations, statistics, product specifications and legal clauses are easy to fake convincingly.
• Your question contains a false premise. "Why did the company cancel the product last year?" invites a reason even if nothing was canceled.
• The information is recent, local or private. Training data has a cutoff, and your own documents were never part of it.
• You made "I do not know" an unwelcome answer. If every question demands an answer, a guess is the path of least resistance.

Ground the answer in text you paste

The most effective technique is to hand over the source and restrict the assistant to it. Reading a supplied text is a much more reliable task than recalling a specific fact from memory.

Answer using only the text between the tags. If the answer is not in the text, write "Not in the source." For every point, include a short verbatim quote from the text in quotation marks.

<source>
[paste the policy, article, contract section or notes here]
</source>

Question: [your question]

Compare two ways to ask about a refund window. The first, "What is the refund policy for this blender?", asks the assistant to recall a specific, changeable fact. It may hand you a plausible number it has no basis for. The second pastes the policy text and demands a quoted sentence for each claim. You still glance at the quotes to confirm they match the page, but that takes seconds, not a research project.

Whenever a real document exists, paste it instead of asking the assistant to remember it.

Ask for uncertainty, not just answers

Give the assistant explicit permission to not know.

If you do not know something or cannot verify it, say so plainly instead of guessing. I would rather have "I am not sure" than a plausible guess. At the end, list what you could not determine and what I should check.

You can also ask it to rate its own claims.

For each factual claim in your answer, label your confidence as High, Medium or Low and say why. Then list what you cannot know from this conversation, such as current prices, recent events, local rules or my private data.

Treat those labels as pointers for your checking, not as measurements. An assistant can be confidently wrong, so a "High" label never replaces verification.

One more habit helps: rewrite leading questions as open ones. Instead of "Why did the company cancel the product?", ask "Did the company cancel the product? If you are not sure or this may be outdated, say so."

Turn its answer into a checking list

Before you rely on an answer, make the assistant list exactly what needs checking.

List every specific, checkable claim in your answer (names, numbers, dates, quotes, citations, legal or technical statements) as a table with the columns Claim, Type, and How I can verify it (what kind of source to check). Do not add new claims.

This takes a vague feeling of "I should double-check this" and converts it into a short task list. The assistant is not verifying anything here. It is helping you see what you would need to verify.

A practical checklist for names, numbers, quotes and citations

Use this before you send, publish, submit or act on anything the assistant wrote.

1. Names. Do the people, companies, products and places exist, and are they spelled correctly? Search the name on its own.
2. Numbers and statistics. Trace each figure to an original source with the same year and units. If you cannot find one, remove the number or describe it as an estimate.
3. Dates and timelines. Confirm against a reliable record. Dates are a common place for small, plausible errors.
4. Quotes. Find the exact words in the original. If you cannot, drop the quotation marks or drop the quote.
5. Citations, titles, authors and links. Open them. Fabricated citations often look real: a believable author, a believable journal, a title that fits. If a link or reference does not lead where it should, do not use it.
6. Laws, rules and regulations. Check the official source for your location and today's date. For anything consequential, ask a qualified professional.
7. Medical, financial and safety claims. Treat the output as a draft for a conversation with a qualified professional or official guidance, not as advice.
8. Technical details. Do the functions, settings, commands and options actually exist in the current documentation?
9. Internal consistency. Do the figures add up? Does the answer contradict itself between paragraphs?
10. Claims about your own material. Does the answer match what you gave it, or did it add things you never said?

How much checking is enough?

Scale the effort to the stakes. This is a rule of thumb, not a formula.

Situation	Reasonable checking
Casual brainstorming	Little or none
An email to a colleague	Names, dates, any numbers
A client proposal or public post	Most of the list
Anything involving health, money, legal obligations or safety	Everything that applies, plus a qualified human

A useful way to calibrate yourself: ask the assistant a factual question in a field you know well, run the checking-list prompt, and verify three claims. Note how many held up. That tells you how much to trust it in areas where you cannot judge as easily.

If an answer is simply not what you wanted, rather than wrong, the companion article on fixes for ChatGPT not giving you what you want covers that separate problem.

Where to go from here

The Prompt Bible has a full chapter on reducing made-up answers, including grounded-versus-ungrounded comparisons, a longer verification checklist and patterns for working with your own documents. If you are earlier in your AI journey and want a plain-English orientation to what these tools can and cannot do, How to Use AI: The Complete Plain-English Guide is the gentler starting point.

Frequently asked questions

Why does ChatGPT make up sources and citations?

It generates text that looks like a citation because that fits the pattern of an answer that has sources. It is not checking a library. Always open the link or look up the title before you use a reference.

Does pasting my own text stop the AI from making things up?

It lowers the risk a great deal, because the task becomes reading rather than remembering. It does not remove it, so ask for quoted evidence and spot-check that the quotes match your text.

Can I just ask the AI whether its answer is correct?

Asking it to flag uncertain claims can help you decide what to check. But it can be confidently wrong about its own answers, so it is not a substitute for checking an outside source.

When is it safe to skip verification?

For low-stakes work like brainstorming names or drafting a casual note, light checking is fine. For anything health, legal, financial or safety related, verify with a qualified professional.