How to Tell If an AI Answer Is Hallucinated
AI hallucinations rarely look like errors. They look confident, specific, and well-formatted. Here is how to catch them before they end up in your product, your pitch deck, or your client's inbox.
An AI answer is probably hallucinated when it hands you a specific fact, a number, a quote, or a citation you cannot independently confirm anywhere else. The tone gives you nothing to go on: fabricated text reads exactly as confident as accurate text, because the model is not tagging one sentence as known and another as guessed. It produces the most statistically likely next words either way. So the fix is to stop reading for confidence and start reading for the parts of an answer that are actually checkable, then check them.
This post is a working detection kit, not a definition. If you want the mechanics of why models fabricate details in the first place, we cover that separately in what an AI hallucination actually is. Here, you get the red flags to scan for, a worked example of catching a fabricated citation, and a verification checklist you can run on any answer before it goes into a deliverable, a client email, or your own product.
The red flags that show up before you verify anything
None of these prove an answer is wrong on their own. Together, they tell you which claims deserve a closer look before anything else.
Suspiciously precise numbers with no visible source. A stat like "34.7% of small businesses" with no report, survey, or link behind it. Real data sources are almost always named; invented ones tend to have oddly exact decimals and nothing pointing to where they came from.
Citations that don't resolve. A case name, DOI, paper title, or page number that returns nothing when you search it, or returns a real source that says something different from what's being quoted.
Confident tone on niche, recent, or long-tail facts. Small companies, minor API changes, recent regulations, obscure case law, anything past the model's training cutoff. Thin training data plus a model built to always produce an answer is a reliable hallucination generator.
Quotes attributed to real people you can't find anywhere else. If a named person supposedly said something and a search turns up nothing verbatim, the quote was likely assembled to sound plausible, not retrieved.
Zero hedging on a topic that should have some. A well-calibrated answer to an obscure question includes some "I'm not certain" or "this may have changed." An answer that's uniformly confident across easy and hard claims is a sign the model isn't distinguishing between them either.
The facts shift when you re-ask. Same question, new session, different number or different name. Consistency isn't proof of accuracy, but inconsistency is close to proof of fabrication.
None of this is about the model being untrustworthy in general. OpenAI's research on why language models hallucinate makes the underlying mechanism plain: standard training and evaluation reward a confident guess over an honest "I don't know," the same way a multiple-choice test rewards guessing over leaving an answer blank. The model isn't lying. It's optimized to never say nothing.
A worked example: catching a fabricated citation
The exchange below is illustrative. We built it to show the pattern clearly, not as a transcript of one specific incident, and the organization and report name in it are invented for the example.
The prompt: "What percentage of small business owners use AI chatbots for customer support?"
The answer: "According to the 2024 Small Business Technology Adoption Report published by the National Small Business Support Council, 34.7% of small business owners now use AI-powered chatbots for customer support, up from 12% in 2022 (page 47)."
That answer has four checkable pieces: an organization, a report title, a specific percentage, and a page number. Here's how each one falls apart on inspection.
Search the organization name. "National Small Business Support Council" returns nothing matching that exact name. Similar-sounding real organizations exist, but none published a report with this title.
Search the exact figure in quotes. "34.7% of small business owners" turns up no independent source. If anything surfaces, it's other AI-generated pages repeating the same unsourced number, not a primary source.
Ask the model for a direct link, then open it. Asked to cite its source, the model produces a URL. It either 404s, redirects to an unrelated page, or points to a real domain that has never published anything with that title.
Try to locate page 47. Since the report doesn't exist, the page number is fabricated by definition. This step is really just confirmation of the first three.
Total time to catch it: a few minutes. Total time it would have cost to publish it, then correct it after a reader or a client's compliance team caught it instead: considerably longer, and worse for your credibility. The pattern generalizes past statistics: fake court citations in legal filings and invented academic references in research summaries follow the same shape, a specific-sounding source that nobody can actually open.
The verification workflow
Run this on any AI answer before you rely on it, especially anything with numbers, names, quotes, or claims about a specific source.
Pull out every checkable claim. List the names, numbers, dates, quotes, and citations separately from the surrounding prose. You're about to verify each one on its own, not the paragraph as a whole.
Search the exact number or quote in quotation marks. An exact-match search is more useful than a paraphrased one. If nothing independent turns up, or only other AI-generated text repeats the same figure, treat it as unverified.
Trace every citation to a source you can actually open. Not a search result snippet, the real document, at the page or section cited. A citation the model produced is a claim to check, not evidence in itself.
Cross-check against a primary source, not a second AI summary. Official documentation, filed data, the organization's own site, a published paper. Two AI tools agreeing with each other isn't independent confirmation; they may share the same gap in training data.
Re-ask in a fresh session and compare. Ask the same question again, ideally a few hours later or in a new conversation. A figure or name that changes between runs is a strong signal the first one wasn't retrieved from anywhere real.
Weight scrutiny by how niche or recent the topic is. Mainstream, well-documented facts are lower risk. Small companies, recent events, niche technical details, and anything past the model's knowledge cutoff get mandatory verification, no exceptions.
When you can't verify it, say so or cut it. Downgrade the claim to "unverified" in your own draft, or remove it. Publishing a plausible-sounding number is a worse outcome than publishing a shorter piece without it.
Why this is worth the extra minutes
Specialized tools with retrieval and citations are not immune. A widely cited Stanford RegLab study of hallucination rates in AI legal research tools found that leading AI legal research products, built specifically to ground answers in real case law, still produced hallucinated or unsupported statements between roughly 17% and 33% of the time, despite marketing language describing them as largely hallucination-free. If tools purpose-built for citation accuracy still miss at that rate, a general-purpose assistant asked an off-the-cuff question deserves at least as much skepticism.
The stakes scale with what you're doing with the answer. A wrong fact in a casual chat is a shrug. A wrong fact in a blog post, a client deliverable, a legal brief, or a product's user-facing copy is a correction, and sometimes worse. Understanding who is on the hook when AI makes a mistake is worth reading before you decide how much verification a given piece of work actually needs; the answer is usually "you," not the model provider.
Two things reduce how often you'll need this workflow in the first place. Vendors sometimes claim their product doesn't hallucinate at all; treat that claim the same way you'd treat any other unverified stat, and read our notes on vetting an AI vendor's accuracy claims before you take it at face value. And on your own prompts, giving the model better context up front about your business, your data, and what "correct" looks like narrows the gap the model would otherwise fill with a guess. Neither replaces verification. Both lower how often you'll need it. Fabricated facts are one entry in a longer list of the broader risks of building with AI, and the same instinct, checking before you trust, applies to most of the others.
Frequently asked questions
Can you tell if an AI answer is hallucinated just by how it sounds?
Not reliably. Confident, fluent phrasing is not evidence of accuracy. Models write hedged and unhedged sentences with the same tone, so you have to check the claim itself, not the delivery.
Are some topics more likely to produce hallucinated answers?
Yes. Narrow, recent, or low-traffic topics are higher risk, because the model has less training data to draw on and fills gaps with plausible-sounding guesses. Case law, niche APIs, small companies, and anything after the model's training cutoff are common trouble spots.
Does asking the AI to cite its sources fix the problem?
No. A model asked for a source can generate a citation that looks real, complete with author names and page numbers, without having retrieved anything. Treat every citation as a claim to verify, not as proof.
What's the fastest single check for a suspicious AI answer?
Search for the specific number, name, or quote in quotation marks in a regular search engine. If nothing independent turns up, or only the AI's own output turns up, treat the claim as unverified.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
