You stop ChatGPT from inventing facts by giving it the source text you want it to use, telling it to use only that text, and then checking every number, name, and quote against the source before you publish.
The model does not lie on purpose — it predicts the most likely next words, and when it has no source, the most likely words are often plausible-sounding but wrong. So the fix is not a magic prompt. The fix is controlling what information the model is allowed to draw from.
Why it happens
ChatGPT was trained to produce fluent text, not to verify truth. When you ask "what percentage of small businesses use AI?" it does not look anything up.
It generates a sentence shaped like an answer, and a number that fits the shape. That is why the invented figure is usually in the right ballpark and completely unsupported. The failure gets worse with specifics: exact dates, study names, page numbers, and quotes are the most likely things to be fabricated, because they are the details a real answer would contain and the model has learned their pattern without the underlying record.
There is a second trap. Asking the model to "be accurate" or "only state facts" does almost nothing. Those instructions describe a goal, not a limit. The model has no way to check itself against a source it was never given. What works is removing the option to invent: hand it the material and forbid anything outside it.
The method that actually works
Paste your source text into the conversation, then add three constraints. First: "Use only the information in the text above." Second: "If something is not in the text, write 'not stated' instead of guessing." Third: "After each factual sentence, note which part of the text it came from." That third instruction is the important one, because it forces the model to point at evidence rather than generate around it. When it cannot point, it usually stops.
Here is a worked example. Suppose you are writing about AI adoption and you ask: "How many AI tools does the average marketing team use?" A model with no source might answer with a specific number and even name a survey.
You cannot verify either. Now paste in your own verified material — for instance, our AI tool database, which holds 360 AI tools with pricing and capability snapshots recorded at verification time, most recently 2026-09-18 — and ask the same question. The honest answer becomes: that count is not stated, so write "not stated." You have lost a sentence and gained a fact you can defend.
A useful tip that goes beyond the obvious: run the draft through the model a second time with the instruction "list every factual claim in this draft that is not supported by the source text above." Models are noticeably better at spotting unsupported claims than at avoiding them in the first place. Treat that pass as a filter, not a guarantee — you still check the numbers yourself.
Where this breaks down
The method depends entirely on having source text. If you are writing from your own knowledge, there is nothing to constrain the model to, and no prompt will fix that. It also costs time: pasting sources and running a verification pass roughly doubles the work of producing a short piece. For a 200-word social caption, that may not be worth it. For anything with a statistic, a price, or a named person in it, it is.
One more limit worth naming. Even with a source, the model can misread it — flipping a "before" and "after," or attaching a number to the wrong company. Citation-checking catches invented facts, not misread ones. That is why the final step is always a human reading the source and the sentence side by side. Tool choice matters far less than input structure here; a well-sourced prompt in a free chatbot beats a vague prompt in an expensive one.