How to Reduce AI Hallucinations in Your Prompts

TL;DR: To reduce AI hallucinations in your prompts, stack four things: give the model explicit permission to say it doesn't know, ground its claims in quotes from a document you supplied, force it to verify or retract each claim before you see it, and treat any technique that promises zero hallucinations as overselling. None of this teaches the model facts it never had.

Why can't a prompt fix this completely?

Because a hallucination isn't a bug the prompt is stepping around, it's the model doing exactly what its training rewarded: producing a fluent, plausible continuation whether or not that continuation is true. A prompt can change the odds of that continuation being an honest gap instead of a confident guess. It cannot hand the model a fact it was never trained on. If you want the mechanism behind why this happens in the first place, why does ChatGPT make things up covers the training-level cause in depth; this piece stays on what you can actually do about it from the prompt box.

That framing matters because it sets a ceiling on what any of the five techniques below can promise. Each one works by changing an incentive inside a single conversation, either making abstention cheaper, making a claim easier to check, or catching an inconsistency after the fact. None of them add a fact to the model's weights. If a question genuinely has no answer the model was ever exposed to, the best any prompt does is get it to say so instead of guessing, which is a real improvement and a bounded one. Keep that ceiling in mind as you read the techniques, because the marketing around this topic routinely implies a higher one.

One more thing worth saying before the list: none of what follows is tied to a single vendor's API. The techniques below are documented in Anthropic's own guidance and in a Meta AI research paper, but the shapes themselves, permission to abstain, quote extraction, forced retraction, independent verification, repeated sampling, are just instructions written in plain text. They work the same way in a prompt sent to GPT, Gemini, or any other chat model, because none of them depend on a model-specific parameter. What changes between providers is how well the model follows the instruction, not whether the instruction is available to give.

Technique 1: Give the model explicit permission to abstain

The simplest lever, and the one most people skip because it feels like giving the model an excuse. Anthropic's own guidance on this, aimed at Claude but applicable to any chat model, states the technique in one line: "Allow Claude to say 'I don't know': Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information."

The reason this works is almost mechanical. Without that permission, the model is implicitly being scored the way most benchmarks score it, where a wrong guess and a right answer both count as "an answer" and only an explicit abstention counts as a miss. Telling it, in the prompt, that "I don't have enough information" is an acceptable and preferred output removes the pressure to fill a gap with something plausible-sounding.

It feels counterintuitive the first time you write it, because it reads like you're inviting the model to give up. In practice it's the opposite: you're removing the one incentive that makes guessing look like the safer choice. A model with no stated permission to abstain has to weigh "say something plausible" against "admit I don't know" with no signal about which one you actually want, and it defaults toward the answer that looks more helpful on the surface. Naming the trade-off explicitly, and saying which side of it you want, is doing the model's risk calculation for it instead of leaving it to guess what you'd prefer.

Review the attached vendor contract and flag any auto-renewal
clause, termination notice period, and liability cap.

If any of these three items is not clearly stated in the
document, say exactly: "Not specified in this document" for
that item. Do not infer a typical or industry-standard value.

Technique 2: Ground every claim in a quote from your source

If you've supplied a document, contract, transcript, or dataset, the highest-leverage move is asking the model to extract the exact supporting text before it answers, rather than asking it to answer and hoping it stayed inside the document. Anthropic frames this specifically as a long-document technique: "For tasks involving long documents (>20k tokens), ask Claude to extract word-for-word quotes first before performing its task. This grounds its responses in the actual text, reducing hallucinations."

The same guidance names the companion move for cases where you don't want the model reaching for anything outside what you gave it at all: "External knowledge restriction: Explicitly instruct Claude to only use information from provided documents and not its general knowledge." Pair both and you get a two-stage prompt: quote, then answer only from the quotes.

Step 1: Read the attached policy document. Extract every
sentence that is directly relevant to data retention periods.
Number each extracted sentence.

Step 2: Using ONLY the numbered quotes from Step 1, answer:
what is the maximum data retention period stated in this
policy? Cite the quote number your answer rests on. If no
quote addresses retention period, say so explicitly instead
of estimating.

Technique 3: Force a citation, then force a retraction

Asking for a citation on its own is a weaker guard than it looks like, because generating a citation-shaped string is no harder for a model than generating any other fluent sentence: plausible author, plausible year, plausible source, none of it checked against anything real. The fix Anthropic documents is a second pass specifically aimed at that gap: "Make Claude's response auditable by having it cite quotes and sources for each of its claims. You can also have Claude verify each claim by finding a supporting quote after it generates a response. If it can't find a quote, it must retract the claim."

That second sentence is the part worth keeping. The citation request alone catches nothing; the mandatory retraction step is what does the work, because it changes the cost of an unsupported claim from "invisible" to "must be visibly deleted."

Draft a summary of the attached research briefs for our
investor update, citing the source document for each claim.

Then, review your own draft. For every claim, locate the exact
sentence in the source that supports it. If you cannot find
one, delete the claim and mark the deletion with [REMOVED -
unsupported].

Technique 4: Chain-of-Verification, for when there's no document to ground against

Everything above assumes you handed the model a source. Closed-book questions, the ones with no document attached at all, need a different structure, and there's a specific, named, published technique for this rather than a vague "double-check yourself" instruction.

Meta AI's Chain-of-Verification (CoVe) paper, published on arXiv in September 2023 by Dhuliawala and colleagues, defines a four-step method: the model "first (i) drafts an initial response; then (ii) plans verification questions to fact-check its draft; (iii) answers those questions independently so the answers are not biased by other responses; and (iv) generates its final verified response." The independence of step (iii) is the load-bearing detail, the model answers its own fact-check questions without seeing its first draft, so it can't just confirm whatever it already said.

Skip that independence and the whole method collapses into theater. Ask the model to "check your answer above" in a single follow-up turn and it can see its own draft while checking it, which means the same pattern-completion pressure that produced the wrong claim in the first place is sitting right there in context, ready to rubber-stamp it. Forcing the verification questions to be answered blind, as their own separate step with no view of the draft, is what makes this meaningfully different from just asking "are you sure?" twice.

Step 1: Draft an answer to my question.

Step 2: Write 3-5 specific fact-checking questions that would
verify the individual claims in your draft. List only the
questions, no answers yet.

Step 3: Answer each verification question independently, as if
you had not seen your Step 1 draft.

Step 4: Compare your Step 3 answers against your Step 1 draft.
Revise the draft to match the verified answers, and remove
anything the verification contradicted.

The paper reports a real result worth quoting precisely, and worth reading with its scope intact rather than generalized past what it tested. On a longform biography-generation benchmark, scored with the FactScore metric, a Llama 65B few-shot baseline scored 55.9; the full "CoVe (factor+revise)" variant scored 71.4 on the same base model. That's a meaningful gain, on a paper worth building the technique from. It is also a 2023 result on Llama 65B, an open-weight base model that predates the current generation of frontier chat models by a wide margin. Nothing in that paper tells you what the same four steps do to a 2026 frontier model's hallucination rate, and this piece won't imply otherwise by omission. Try it, and judge it against your own task rather than importing the paper's number as your own.

Technique 5: Best-of-N, for catching inconsistency cheaply

Anthropic's guidance also names a lower-effort check: "Run Claude through the same prompt multiple times and compare the outputs. Inconsistencies across outputs could indicate hallucinations." This one is worth being precise about what it catches and what it doesn't. If a model states a specific fact differently across three independent calls, at least two of those answers are wrong, and you've caught that cheaply. If it states the same wrong fact all three times with equal confidence, best-of-N shows you three identical, equally wrong answers and nothing else. It's a consistency check, not a truth check, and the distinction matters for what you use it on.

What if you don't have a source document to ground against at all?

Everything in Technique 2 assumes you can hand the model a document. When the question is genuinely open (current events, a fact you don't have on hand, anything time-sensitive) grounding means turning on retrieval instead of pasting a file, and that comes with its own prompt shape. Prompting ChatGPT search covers when to reach for browsing versus a closed-book prompt, and how to phrase the query so the model actually uses what it retrieves instead of blending it with what it already "remembers."

No row in that table reads "catches everything." That's deliberate: the honest answer to "which one works" is that none of them alone does, and a confidently, consistently wrong answer is the failure mode every technique here struggles with most, because consistency is exactly what makes it look trustworthy.

Stacking these five techniques moves your odds, not your certainty. If the model keeps hedging every sentence with a disclaimer even after you've asked it to reduce fabricated claims, why does ChatGPT keep apologising is the companion piece for the opposite-looking problem that shares some of the same root cause. Treat every one of the prompts above as a way to lower the rate of a specific failure, verify the output that matters most to you regardless, and don't let any tool, including this one, tell you it's solved.