Table of Contents
TL;DR: AI hallucination is the technical term for what happens when an AI invents facts and presents them with full confidence. It cites studies that don’t exist, attributes quotes to the wrong person, gets dates wrong by decades, and writes all of it in clean, authoritative prose. It happens on every major model, multiple times per long document, and it cannot be eliminated by better prompting. The fix is to verify every specific claim before you put your name on it. This guide tells you what hallucination looks like, why it happens, where it shows up most, and how to catch it before it catches you.
This guide is for anyone publishing under their own name who is using AI as part of the process. Writers, consultants, researchers, executives writing books or articles, attorneys drafting briefs, anyone whose reputation rides on whether the facts in their work are real. The questions are the ones I get asked most. The answers are short on purpose.
What is an AI hallucination?
A hallucination is any output where an AI states something as fact that isn’t true, with full confidence and no signal to the user that the system isn’t sure. Hallucinations aren’t typos or weird phrasing. They are invented citations, fabricated quotes, wrong dates, nonexistent studies, made-up legal cases, plausible-sounding statistics with no source. The output reads completely normal. The fact it states is just not real.
Why does AI hallucinate?
Large language models work by predicting the next likely word given everything that came before. They are not databases. They have no concept of “I know this” versus “I don’t know this.” When a model is asked for a fact it doesn’t have, it generates the most plausible-sounding word, then the next, then the next. The result is a fluent sentence that may or may not be true, and the model has no internal signal that distinguishes the two.
It’s not lying. It’s not broken. It’s doing exactly what it was built to do, which is to produce text that sounds right. Whether the text is right is a separate question the model wasn’t built to answer.
How often does it happen?
Often. Multiple independent studies have measured it. Citation invention in legal briefs has been documented in published court cases. Statistical invention happens in nearly every long document where the model is asked for specific numbers it wasn’t given. The rate varies by model, by domain, and by how confidently you push the model into territory it doesn’t know, but the baseline is high enough that you should assume any specific claim in AI output may be invented until you verify it.
What kinds of facts does it invent most?
The pattern is consistent across models. The most common categories:
- ► Academic citations. Authors, journals, page numbers, study titles, all invented with the right shape and zero existence.
- ► Quotes attributed to real people. The person exists. The quote does not.
- ► Statistics with specific decimal places. The more precise the number sounds, the more confident the invention.
- ► Dates of events. Off by months, years, sometimes decades.
- ► Legal cases. Plaintiff, defendant, court, year, the entire citation fabricated.
- ► Book titles and authors that don’t exist.
- ► Technical specifications, especially for older or less common products.
How do I spot a hallucination in my own draft?
Read every specific claim and ask one question: where did this come from?
If the answer is “the AI said it” and there’s no source the AI gave you that you can independently verify, treat the claim as unverified. That means every name, every date, every number, every quote, every citation. Especially the ones that sound the most authoritative, because the model produces its most confident invention precisely when it has the least to work from.
Specific tells worth scanning for:
- ► Citations that name a researcher and a paper but no DOI or URL.
- ► Statistics rounded to weird decimal places (43.7%, 28.3%) that you can’t trace to a source.
- ► Quotes that sound exactly like what the model would have written, attributed to someone famous.
- ► Dates that don’t match what you can find elsewhere.
- ► Any reference to a “recent study” with no specific journal or year.
Will turning on web search fix it?
Partially. When the model has live search access, it can pull actual sources and cite them, and the citation invention rate drops significantly. The problem is that the model still synthesizes the result, and the synthesis can still misrepresent what the sources said. So you trade citation invention for paraphrasing errors, which are harder to catch because the sources are real but what the model says they say isn’t quite accurate.
Web search is an improvement, not a fix. The verification step is still on you.
Can I just ask the AI to fact-check itself?
No. Asking the model “are you sure?” or “is this true?” doesn’t pull from a different knowledge base. It just generates a new response using the same mechanism. Often the model will confidently confirm an invention because the invention sounds plausible to it on a second read.
Some models will flag low confidence on certain outputs, and those flags are worth paying attention to. The absence of a flag is not a confirmation of accuracy.
What’s the fix?
Verify every specific claim before publication. There is no shortcut. The workflow that works:
- ► Generate the draft with AI if that’s part of your process, knowing the draft is a starting point not an authority.
- ► Highlight every specific factual claim: names, dates, numbers, quotes, citations.
- ► For each one, find an independent source. Not the AI. The original.
- ► If you can’t verify a claim, cut it or replace it with something you can verify.
- ► Ship the verified version.
This is slower than just publishing the AI draft. It’s also the only thing that prevents you from publishing a hallucination under your own name.
What if I publish a hallucination by accident?
Fix it publicly and quickly. The standard for handling published mistakes is the same as it was before AI existed: name the error, correct it, explain how it happened. The reputation damage from a corrected mistake is small. The reputation damage from a hallucination that stays uncorrected on your byline is permanent.
Attorneys have been sanctioned for citing AI-invented cases in court filings. Authors have had books withdrawn from publication for fabricated citations the AI introduced. Journalists have lost careers over invented quotes. The cost of catching it before publication is always less than the cost of catching it after.
Checklist before you publish anything AI helped you write
- ► Every citation traced to an independent source you actually clicked on.
- ► Every quote checked against the speaker’s actual published or recorded statements.
- ► Every statistic verified against the original study, not a secondary mention.
- ► Every date confirmed against a primary source.
- ► Every name spelled the way it’s actually spelled.
- ► Every legal case verified in a legal database, not just confirmed by a second AI prompt.
If you can’t check the box, cut the claim. There is no version of “I think the AI got this one right” that’s worth the risk.
If you want the broader case for why your work has to clear this bar, the piece on why AI alone isn’t the problem walks through what readers actually catch and how they catch it.
Frequently Asked Questions