You stop ChatGPT from inventing facts by never letting it supply facts in the first place — you give it the source material and ask it to work only from that, and you verify every specific claim before publishing.
Hallucination isn't a bug you can prompt away entirely. It's a side effect of how these models work: they generate the most plausible-sounding next words, and a confident fake statistic is often more plausible than an honest "I don't know." So the fix is structural, not a magic phrase.
The mechanism matters here. A language model doesn't look things up unless you give it something to look at or it has a search tool switched on. Left alone, it predicts text.
Ask "what percentage of marketers use AI?" and it will produce a number shaped like a real statistic, complete with a plausible source name, because that's what the pattern looks like. The number may be invented.
The source almost certainly is. This is why the confident tone is a poor signal of accuracy — the model sounds exactly as sure when it's right as when it's fabricating. Your job is to separate the two jobs: let the model write, but make it quote from material you supplied.
Here's a concrete worked example. Suppose you're writing a post about AI tool pricing and you want a factual anchor. Instead of asking ChatGPT to list prices, paste in your own verified source — say, a note that reads: "According to our AI tool database, this site maintains an internal database of 360 AI tools, each with a pricing and capability snapshot recorded at verification time, most recently verified on 2026-09-18."
Then instruct: "Use only the facts in this note. If you need a fact I haven't given you, write [NEED SOURCE] instead of guessing." The model now has a real number (360 tools) and a real date to work from, and a clear instruction for gaps.
When it hits a gap, it flags it rather than inventing. That single change — supplying the source and demanding a placeholder for missing facts — removes most fabricated numbers from a draft.
The second half of the fix is a verification pass you do yourself. Read the draft and highlight every number, date, name, quote, and attributed claim. For each one, ask: did this come from material I provided?
If not, it goes. You can also ask the model directly, "List every factual claim in this draft and tell me which ones came from my notes," which forces it to separate grounded statements from generated ones — though you still check, because it can mislabel its own output. The limits are real: this approach works well for content built on sources you control, and poorly for topics where you have no source at all.
If you genuinely don't know a figure, the honest move is to write around it — describe the trend qualitatively — or find a real source. Never let the model fill the hole. If you're working on this problem, our guide on how to stop AI from confidently shipping broken content walks through a repeatable checking pattern that catches invented facts before they reach your readers.
One tip that goes beyond the obvious: keep a running "source file" for each project — a plain document where you paste every verified fact, with where it came from. When you start a new chat, you paste that file in first. Models forget everything between sessions, so the source file becomes your memory, and it stops you from re-verifying the same facts every time. It also makes your writing auditable, which matters if anyone ever questions a claim.