Why does ChatGPT make things up in the first place?

It predicts the next likely word based on patterns, not a database of verified facts. When it does not know something, it does not stop. It generates the most plausible-sounding answer, which can be wrong. That is why it states fake citations and invented statistics with the same confidence as real ones. The fix is to give it the facts and check the rest.

What is the single most effective way to reduce hallucinations?

Grounding. Paste the actual source text into the chat and tell the model to answer only from what you provided. Add a line like "if the answer is not in the text, say you do not know." This turns the model from a guesser into a reader, and it removes most made-up answers in one move.

Can I trust the links and sources ChatGPT gives me?

No, not without checking. Models are known to invent URLs, paper titles, case numbers, and author names that look completely real. Always open the link yourself and confirm the source exists and actually says what was claimed. A citation you have not clicked is not a citation.

Does saying "do not hallucinate" in the prompt work?

Not on its own. The model cannot tell when it is guessing, so an instruction to "be accurate" or "do not make things up" does little by itself. What works is removing the need to guess: give it the source, tell it to say "I do not know" when the answer is missing, and then verify the specifics yourself.

GuidesFoundations

1 / 10

Start herebeginner5 min read

How to Stop ChatGPT Hallucinations and Fact-Check Its Answers

By Automate Basics

Cut hallucinations by grounding the model. Paste the source text and add "only use what I gave you; if it is not there, say you do not know and do not guess." Then verify any name, number, quote, or citation yourself before you send it. ChatGPT predicts likely words, not true ones, so grounding plus a quick check is what makes its output safe to use at work.

Swipe up to begin

Concept

Why it invents things so confidently

ChatGPT does not look up facts. It predicts the next likely word from patterns in its training. Most of the time that lands on something true. When it does not know, it does not stop or warn you. It fills the gap with the most plausible-sounding text, stated with the exact same confidence as a real answer. That is why it produces fake citations, invented statistics, and made-up quotes that read perfectly. Once you accept that the model is a fluent guesser, not a fact database, the rest of this makes sense.

Step 1 of 4

Step 1: Ground it in your own sources

The biggest fix is to stop asking the model to recall and start asking it to read. Paste the actual text, report, email thread, or policy into the chat. Then ask your question about that text. A model answering from material in front of it makes up far less than one answering from memory. So instead of "what is our refund window," paste the refund policy and ask "according to this, what is our refund window?" You just turned a guess into a lookup.

Step 2 of 4

Step 2: Add the one line that gives it permission to not know

Most hallucinations happen because the model thinks it must produce an answer. Remove that pressure. Add this to your prompt: "Only use the text I gave you. If the answer is not in there, say you do not know and do not guess." That single instruction lets the model return "the document does not cover this" instead of inventing a clause. Pair it with grounding from Step 1 and you have shut down the most common failure: a confident answer to a question your source never addressed.

Step 3 of 4

Step 3: Verify every name, number, quote, and citation

Grounding cuts the risk. It does not remove it. Before anything leaves your hands, check the specifics the model is worst at:

Names and titles of people, companies, and products.
Numbers: dates, prices, percentages, statistics.
Direct quotes, especially ones in quotation marks.
Citations: links, sources, case numbers, study titles.

These four are where fabrication hides. If you pasted the source, confirm each one traces back to it. If you did not, look it up. A claim you have not checked is not a fact yet.

Step 4 of 4

Step 4: Never trust a link you have not opened

This one deserves its own rule because it burns people in public. Models invent URLs, article titles, author names, legal cases, and DOI numbers that look completely real. A lawyer was sanctioned for filing court citations ChatGPT made up. The fix takes ten seconds: click every link the model gives you and confirm the page exists and says what was claimed. If a "source" cannot be opened and read, treat it as fiction until proven otherwise. Do not paste a citation into your work on the model's word alone.

Concept

Make the model check its own work

You can get a second pass for free. After an answer, ask the model to audit it: "Go through your answer and mark each claim as either supported by the text I gave you or not in the text." It will often flag its own weak spots and pull back claims it cannot ground. Another useful move: ask the same question in a fresh chat, or in a second tool like Claude or Gemini. If two independent answers disagree on a fact, that fact is exactly the one you need to verify yourself.

Concept

Match the tool to how much you trust the output

A quick brainstorm and a customer-facing number carry different stakes, so pick accordingly:

Low stakes (ideas, rough drafts, rewording): use any chat freely, light checking.
Facts that matter (sent to a client, a boss, the public): ground it, then verify every specific.
"What does our document actually say": use a citation-first tool like NotebookLM, which answers only from your files and links each claim to the source page.

The skill is not avoiding AI. It is knowing which answers you can ship as-is and which ones you check first.

Concept

Try this on your next answer

Take the next thing you ask ChatGPT for work. Before you send its reply anywhere, do three things: paste the real source and ask it to answer only from that, add "say you do not know if it is not in the text," then click every link and confirm every number. Do this five times and it becomes automatic. You will catch the one fabricated detail that would have cost you, and you will stop fearing the tool.

Try this now

Ask the tool something from your own field that you know cold, push it for specifics, and watch where it bends. Ten minutes of that teaches the failure pattern better than any rule list — then ground your next real question in your own sources.

That’s the whole lesson. Try it on a real task while it is fresh, then come back for the next one.

Questions people actually ask

Why does ChatGPT make things up in the first place?: It predicts the next likely word based on patterns, not a database of verified facts. When it does not know something, it does not stop. It generates the most plausible-sounding answer, which can be wrong. That is why it states fake citations and invented statistics with the same confidence as real ones. The fix is to give it the facts and check the rest.
What is the single most effective way to reduce hallucinations?: Grounding. Paste the actual source text into the chat and tell the model to answer only from what you provided. Add a line like "if the answer is not in the text, say you do not know." This turns the model from a guesser into a reader, and it removes most made-up answers in one move.
Can I trust the links and sources ChatGPT gives me?: No, not without checking. Models are known to invent URLs, paper titles, case numbers, and author names that look completely real. Always open the link yourself and confirm the source exists and actually says what was claimed. A citation you have not clicked is not a citation.
Does saying "do not hallucinate" in the prompt work?: Not on its own. The model cannot tell when it is guessing, so an instruction to "be accurate" or "do not make things up" does little by itself. What works is removing the need to guess: give it the source, tell it to say "I do not know" when the answer is missing, and then verify the specifics yourself.

Related guides

The same corner of the library, one job further on.