You ask a simple question and get a confident, detailed answer. Then you check and find that a name, a date, or a source does not exist. This is often called a hallucination: the AI states something that sounds right but is not true.
This guide shows how to stop ChatGPT from hallucinating as much as possible. The prompts here ask for safer answers, help you check the result, and fix wrong replies. They can reduce errors, but they cannot remove them, so you still need to verify what matters.
Use the cheat sheet below as a quick reference. Then read on to see when to use each prompt and how it looks filled in.
How to Use These Prompts
Words in [brackets] are blanks. Replace each one with your own details before you send the prompt.
Before: Use only this text: [paste]
After: Use only this text: The Elm Street Library opens at 9 a.m. on weekdays and closes at 6 p.m.
You can add these lines to almost any request. They work the same way in ChatGPT, Claude, and Gemini. No prompt is a guarantee, so treat every answer as a draft and check important facts in a reliable source.
Ask Safer Questions
Use only this text: [paste]
When the AI works from text you provide, it has less room to invent. Paste the source and tell it to stay inside it. This works well for summaries and questions about a document.
Example: Use only this text: The Elm Street Library opens at 9 a.m. on weekdays and closes at 6 p.m. Answer: When can I visit on a Tuesday?
If unsure, say you’re unsure
By default, the AI tends to give an answer. This line invites it to admit uncertainty instead. It does not make the AI always right, but it can lead to more honest replies.
Example: If unsure, say you're unsure. What year did the Elm Street Library open?
Don’t guess names or numbers
Made-up names, dates, and statistics are a common problem. This line asks the AI to flag missing facts rather than fill in blanks. Check any remaining names and numbers yourself.
Example: Don't guess names or numbers. Write a short intro for a talk about local libraries, and leave a blank where I need to add a figure.
Give a source for each claim
Asking for sources shows you where to look. But the AI can also invent sources, so open each one and confirm it exists and says what is claimed. If you cannot find it, do not use it.
Example: Give a source for each claim in your answer about how libraries lend e-books. Tell me if you are unsure a source exists.
Answer only about [topic]
A narrow scope leaves less room for wandering. Tell the AI exactly what to cover and what to skip. This is useful when you want a short, focused answer.
Example: Answer only about library card requirements. Do not discuss fees or opening hours.
Check the Answer
List claims I should verify
After an answer, ask which statements deserve a second look. The AI may point to dates, numbers, and names. Use the list as a checklist, then verify each item in a reliable source.
Example: List claims I should verify in your last answer about the history of public libraries.
How confident are you, and why?
This prompt asks the AI to describe its certainty. It can show weak spots, but confidence statements can also be wrong. Treat the reply as a hint, not proof.
Example: How confident are you, and why? Explain which parts of your answer are most likely to be wrong.
What could be wrong here?
Asking the AI to challenge its own answer can reveal errors or assumptions. It works best when you give it room to think of risks. Verify anything it flags.
Example: What could be wrong here? Review your answer about how to renew a library book online.
Quote the part that says this
When you give the AI a text, this prompt ties each claim to a line in that text. If it cannot find a quote, the claim may not be supported. Compare the quote with the original yourself.
Example: Quote the part of my text that says the library closes on Sundays. If there is none, say so.
Check this against [source]
Paste a document you trust and ask the AI to compare. It can flag places where the answer and the source disagree. You still need to read both and decide.
Example: Check this summary against the library's policy document below, and list any statements the document does not support.
Fix a Wrong Answer
That seems wrong; recheck
If something looks off, say so. The AI may correct itself, but it may also repeat or change the error. Verify the new answer in a reliable source.
Example: That seems wrong; recheck the date you gave for the library's opening, and explain what you are unsure about.
Show your steps
For math and logic, seeing the steps helps you find the error. Ask the AI to write out each step. Then check them with a calculator or by hand.
Example: Show your steps for how you calculated the total late fees for 12 books at 10 cents a day over 5 days.
Start over using only my facts
When a conversation has gone off track, reset it. Give the facts you know are true, and tell the AI to use only those. This removes earlier guesses from the new answer.
Example: Start over using only my facts: the library opens at 9 a.m., closes at 6 p.m., and is closed on Sundays. Write a short notice for the front door.
Mark anything you invented
Ask the AI to label any detail it made up. It may not catch every guess, so still verify the rest. The label helps you spot places to check first.
Example: Mark anything you invented in this draft with the word GUESS.
Ask a narrower question
Broad questions invite broad, shaky answers. Smaller questions are easier to check. Break a big question into parts and verify each answer.
Example: Ask a narrower question: instead of "Tell me about library history," ask "What is a lending library?"
Common Mistakes
Trusting a confident tone
An answer that sounds sure can still be wrong. Tone is not a measure of truth. Quick fix: ask yourself, “Would I check this if a stranger said it?”
Skipping source checks
If the AI gives a link, book, or quote, open it or look it up. Do this before you share it. Sources that sound real may not exist.
Asking for exact numbers
Statistics, dates, and quotes are common places for errors. Verify them in the original source. If you cannot confirm a number, leave it out.
High-stakes advice
Medical, legal, financial, and safety questions need a qualified professional. Use AI to prepare questions or understand terms, not to make the final call.
Pasting vague questions
A vague question forces the AI to guess what you mean. Add context, such as who the answer is for and what you plan to do with it. Clear questions lead to better answers.
Pro Tip
Treat AI as a first draft. Verify every name, number, date, and quote before you share or publish it, and keep a short list of the claims you confirmed.
Example: Here is my draft. List every name, number, date, and quote in it so I can check each one in a reliable source.
FAQ
Why does ChatGPT make things up?
AI tools generate text that fits patterns in language, and they do not check facts the way a person might. This can produce answers that sound right but are false. That is why checking matters.
Can I stop hallucinations completely?
No. The prompts here can reduce the chance of errors, and providing your own source text helps. You still need to verify important facts.
How do I fact-check AI answers?
Look up key claims in reliable sources, such as official websites, textbooks, or primary documents. Open any cited source and confirm it says what the AI claims. For high-stakes questions, ask a qualified professional.
Final Thoughts
You cannot make an AI tool perfectly accurate, but you can make errors easier to catch. Give it source text, ask it to flag uncertainty, and verify what matters. Keep private information out of the chat.
If this guide helped, save the cheat sheet on Pinterest and bookmark this page for your next fact-check.