How to Stop ChatGPT From Making Things Up
You cannot stop it entirely, and anyone selling you a prompt that does is wrong. A language model generates the most plausible continuation of your text; when it does not have a fact, plausible is all it has, and the output looks identical either way. What you can do is make fabrication much less likely and much easier to catch. This page covers the five things that measurably help, what the research behind each actually claims, and the two-minute check for what still gets through.
Written against published research and provider documentation as of 12 September 2026. Where a technique's evidence is thinner than its reputation, that is stated rather than smoothed over.
Key Takeaways
- Give it the source rather than asking it to recall one. Retrieval-grounded models "generate more specific, diverse and factual language" than models relying on training alone (Lewis et al., NeurIPS 2020, retrieved 2026-09-05, SOURCED).
- Turn on search for anything current. Verifying against a live source is the one intervention whose paper explicitly claims it "overcomes issues of hallucination" (Yao et al., ICLR 2023, retrieved 2026-09-05, SOURCED).
- Make it check its own answer. Chain-of-verification — answer, then generate and answer verification questions — is the one method whose paper is titled for hallucination reduction (Dhuliawala et al., 2023, retrieved 2026-09-05, SOURCED).
- Asking for step-by-step reasoning helps less than people think. That paper measures arithmetic, commonsense and symbolic reasoning — not fabrication (Wei et al., 2022, retrieved 2026-09-05, SOURCED).
- Citations are the highest-risk output. A model will produce a correctly formatted case name, DOI or page number whether or not it exists. Every one needs opening.
Why it happens at all
The model is not looking anything up unless you have given it something to look at. It is predicting text. Where its training gave it a strong, consistent signal — a famous date, a widely documented API — the prediction is usually right. Where the signal is thin, sparse or contested, it still produces fluent output, because fluency is what it optimises.
That is the whole problem in one line: the model has no separate channel for "I do not know". Uncertainty and confidence render identically.
Two consequences follow, and they shape everything below. Fabrication clusters around specifics — names, numbers, dates, citations, prices — rather than general explanation. And it gets worse on obscure questions, which are exactly the ones you needed help with.
For the underlying mechanism, see why AI models hallucinate and what hallucination actually is.
1. Paste the source in
The single most effective change, and the least discussed, because it is not a prompting trick.
If the answer exists in a document, give it the document. Asking a model to summarise a page you have pasted is a fundamentally different task from asking it to recall what a page says: one is reading comprehension, the other is memory.
The research behind retrieval-augmented generation reports that grounded models "generate more specific, diverse and factual language" than a comparable model without retrieval. Note what it does not say — it never uses the word hallucination, and it never claims to be the most effective method available, a superlative widely attributed to it including, until recently, by us.
And it introduces a new failure: retrieving the wrong document. Grounding moves the error from "invented" to "sourced from the wrong place", which is easier to catch but not the same as solved.
Do: paste the actual page, contract, docs or transcript. Do not: name a document and ask what it says.
2. Turn search on for anything that changes
Prices, versions, policies, current officeholders, anything dated. A model without live retrieval answers from training data with a cut-off, and it will not tell you the answer is stale — staleness reads exactly like accuracy.
Of all the techniques here, tool use and live retrieval have the most direct evidence: the ReAct paper states the approach "overcomes issues of hallucination and error propagation prevalent in chain-of-thought reasoning".
The catch: the model can now retrieve a wrong or low-quality page and repeat it with the same confidence. You have changed the failure mode, not removed it. Check what it actually cited.
3. Make it verify its own answer
Chain-of-verification is the most under-used technique on this list. The model answers, then generates verification questions about its own answer, answers those independently, and revises.
You can run it manually in two messages:
"List every factual claim in your previous answer as a separate line, with nothing else."
"For each claim, state how you know it and how confident you are. Mark any you cannot support."
The second prompt is where fabrications surface — a model asked to justify a claim in isolation will frequently drop it. This works because you have separated generation from checking, which is the same reason the published method works.
4. Ask for abstention, and mean it
Telling a model it may say "I don't know" raises how often it does. It is worth doing, and it is the weakest item here.
No prompt gives a model facts it does not have. Instructing it to cite sources makes it produce source-shaped text, which is not the same as producing real sources. Instructing it to be accurate does not make it accurate — there is no accuracy dial being left switched off.
Prompting shifts the distribution of outputs. It does not change what the model knows. Our prompting tips for accuracy covers the phrasing that helps, with that limit stated.
5. Do not rely on step-by-step reasoning to fix facts
"Think step by step" is the most repeated advice in this category and the most over-claimed.
The chain-of-thought paper measures performance on arithmetic, commonsense and symbolic reasoning. Those are reasoning tasks. Better reasoning and fewer fabrications are related, and they are not the same claim — a model can reason impeccably from an invented premise, and step-by-step output can make a fabrication more convincing by dressing it in working.
It genuinely helps for multi-step problems. It is not a factuality fix. See chain-of-thought and hallucination, where the originating papers are cited directly.
The two-minute check for what gets through
None of the above is sufficient, so this is the part that actually protects you.
Open every citation. This is non-negotiable and it is where the worst failures live. A model will produce a case name, a DOI, a page number or a journal reference in perfect format whether or not it exists. Attorneys have been sanctioned over exactly this. Format is not evidence.
Check every number against its source. Not against a summary — against the sentence that contains it. The subtlest error we found in our own audit was a study's speed finding restated as an automation finding: same paper, different claim.
Confirm named things exist. Products, features, plans, people. Type the vendor's own domain. In our audit, five recommended products did not exist at all.
Be suspicious of the specific and obscure. A precise figure on a niche topic is the highest-risk output a model produces, because that is exactly where its training signal was thinnest.
The full nine-step version is in nine checks that catch a fabricated round-up.
What does not work
AI detectors. They answer "was this machine-written", not "is this true", and on their own question the evidence is poor — seven detectors averaged a 61.22% false-positive rate on non-native English writing. See do AI detectors actually work.
Asking the model whether it is sure. Confidence expressed in text is generated text. It is not a measurement of anything internal.
Assuming a newer or larger model has solved it. Hallucination rates vary by model and have improved, but public benchmarks still show most models above a 15% rate on grounded summarisation tasks. It is a spectrum, not a solved problem.
How This Guide Was Sourced
Written and maintained by the LogicBalls editorial team (logicballs.com). Disclosure: LogicBalls builds AI writing tools. A company selling AI writing has an obvious interest in how AI errors are understood, which is why each technique here is tied to the paper that introduced it, including where that paper claims less than the technique's reputation suggests.
Sources. Retrieval-augmented generation: Lewis et al., NeurIPS 2020. Chain-of-thought: Wei et al., 2022. Chain-of-verification: Dhuliawala et al., 2023. ReAct: Yao et al., ICLR 2023. All fetched from arXiv on 2026-09-05 and linked inline.
What could not be fetched. OpenAI's own documentation pages returned HTTP 403 to automated requests during preparation, so no claim here is sourced to OpenAI's published material and none is quoted from it. Statements about ChatGPT's behaviour are drawn from the general research literature on language models rather than from provider documentation, and are described as such.
What is not claimed. We ran no benchmark and publish no measurement of our own. Every quantitative figure above is someone else's, linked and dated.
Model behaviour changes. Retrieval defaults, search availability and error rates shift between releases. Everything here carries a retrieval date for that reason.
No LogicBalls telemetry is used in this guide.
Frequently Asked Questions
Can a prompt stop ChatGPT hallucinating completely?
No. Prompting changes which outputs are more likely; it cannot give the model facts it does not hold. Anyone selling a prompt that "eliminates" hallucination is describing something no published method achieves.
Does turning on web search fix it?
It helps substantially for anything current, and it introduces a different failure: retrieving a wrong or low-quality page and repeating it confidently. Check what it actually cited rather than assuming retrieval means correct.
Why does it invent citations specifically?
Citations have a highly regular format, which is easy to generate convincingly, and the underlying facts are sparse in training data. That combination — easy to imitate, hard to recall — is the worst case. Open every one.
Are newer models meaningfully better?
Measurably, yes, and not enough to skip verification. Public benchmarks still place most models above a 15% hallucination rate on grounded summarisation.
Is this specific to ChatGPT?
No. Every large language model shares the mechanism, and rates differ by model and task. The techniques here apply to all of them.
Conclusion
You cannot stop it. You can make it much rarer — paste the source in, turn search on for anything dated, make the model check its own claims — and you can make what remains easy to catch.
The check that matters most takes two minutes: open every citation, check every number against the sentence it came from, and confirm that named things exist. A model produces a fake reference and a real one in identical format, and that is the whole reason the check cannot be skipped.
