You stop ChatGPT from inventing facts by giving it the facts yourself and forbidding it from adding any others — a technique called grounding, where you paste your source material into the prompt and instruct the model to use only that material.
ChatGPT doesn't look things up unless you tell it to. It predicts what words would plausibly come next, and a confident-sounding statistic is just as plausible to the model as a true one. That's why the fix isn't a better prompt wording, it's a change in who supplies the information. You supply it.
The model's job shrinks to arranging and phrasing what you gave it, which is a task it does reliably.
The mechanism matters because it explains why some anti-hallucination tricks work and others don't. Telling ChatGPT "don't make things up" is close to useless on its own — the model has no way to check itself against reality, so it has nothing to compare its output to. But when you paste in a source, the model can compare its draft against that text, and you've created a closed world it can be held to.
A practical version: open your draft, paste the reference paragraph above the instructions, and add a rule like "Every number, name, and date in your output must appear in the text above. If I ask for something not covered there, reply 'not in the source' instead of guessing." That last clause is the important one. Without it, the model fills gaps silently, because filling gaps is what it's built to do.
Here's a worked example using real material. Suppose you're writing about chatbot pricing and you want a paragraph comparing plans. Paste in your notes: ChatGPT is developed by OpenAI, with a free tier on GPT-4o mini, Plus at $20/mo, and Pro at $200/mo; Claude is developed by Anthropic, with a free tier, Pro at $17/mo annual or $20/mo, and Max from $100/mo.
According to our AI tool database, ChatGPT carries an editorial rating of 4.9/5 and Claude 4.8/5. Now ask for a 100-word comparison. The model can only rearrange those figures.
It cannot invent a "$35 team tier" for either product, because no such number exists in the material it was handed. Compare that to asking the same question cold: you'll often get a confident table with plausible prices that don't match anything, and no warning that they were generated rather than retrieved.
A useful habit that goes beyond the obvious: treat every output as a draft with unverified claims until you've highlighted each number and matched it to your source. This is faster than it sounds. Numbers, dates, and proper nouns are the only things that can be flatly wrong; phrasing is a style question you can fix later.
So do a pass where you read only the digits and names, ignoring the prose. If a figure has no home in your notes, cut it or replace it with a qualitative statement like "pricing varies by tier and changes often." Two limits are worth knowing.
First, grounding only works if your source material is complete — paste in half the facts and the model will still guess at the rest. Second, even a grounded model can misattribute: it might attach the right number to the wrong product. So the final check is always human.
If you want the full checklist for catching this before publishing, see How do I stop AI from confidently shipping broken content (a pattern that actually works)?.