AI Hallucinations Explained: Why AI Makes Things Up (and How to Stop It)

AI hallucinations are no longer a quirk you laugh at in a screenshot. Courts have now logged 2,184 rulings where someone filed AI-invented material, and the count grows by roughly five a day. Yet the same chatbots that invent case law can summarize a document you hand them with under 10% errors. The gap between those two numbers is the whole story, and it's the part most guides skip.

⚡ Short answer: AI hallucinations are confident answers that aren't true: invented facts, fake citations, made-up quotes. They happen because models predict likely words and are trained in ways that reward guessing over saying "I don't know." The single biggest fix is giving the AI the source instead of asking from memory, then checking every name, number and link it returns.

Some links are affiliate links · they support this site at no extra cost to you · Full disclosure

What are AI hallucinations?

Short answer: AI hallucinations are outputs that sound plausible and confident but are false or unsupported. OpenAI's own definition is "plausible but false statements." The model doesn't flag them, so a made-up court case reads exactly like a real one. They're a side effect of how language models work, not a rare bug.

The term AI hallucinations is borrowed loosely from psychology, and it's a slightly misleading one. The AI isn't seeing things. It's filling a gap with the most likely-sounding words, and sometimes the likely-sounding words describe something that never existed.

In practice, AI hallucinations come in four flavors. Knowing which one you're facing tells you how to check it.

Core distinction: a confident tone is a writing style, not evidence. If you want the psychology of why fluent answers feel trustworthy, I cover it in why AI sounds confident when it's wrong.

Why do AI hallucinations happen?

Short answer: Two reasons stack up. Models learn by predicting the next word, so rare facts with no pattern behind them get guessed. Then training and benchmarks grade answers right or wrong, which rewards a bold guess over an honest "I don't know." OpenAI's researchers call it teaching to the test.

The first cause of AI hallucinations is in pretraining. A model sees billions of sentences but no labels saying "this one is false." Spelling and grammar follow patterns, so errors there vanish with scale. A person's birthday or a small company's founding year has no pattern, so the model produces the most plausible-looking guess.

The second cause is newer and more useful to understand. In OpenAI's research on why language models hallucinate, the authors compare it to a multiple-choice exam: leaving a question blank scores zero, a guess sometimes scores a point. Over thousands of questions, the guesser climbs the leaderboard.

Their own numbers make the trade-off visible. On the same fact quiz, one model guessed almost every time and the other often admitted it didn't know:

Read that twice. The model with slightly higher accuracy was wrong three times as often, because it never held back. Accuracy-only scoreboards would rank it higher anyway.

What this means for you: a chatbot's default is to answer. Explicit permission to say "I don't know" changes its incentives in your chat, and it's free. Fix 2 below shows the wording.

How often do AI hallucinations happen?

Short answer: It depends almost entirely on whether the AI has a source in front of it. When summarizing a document they were given, most current models hallucinate roughly 2% to 12% of the time. When answering hard factual questions from memory, rates range from 22% to 94%. Same models, wildly different rates of AI hallucinations.

The low numbers come from the Vectara hallucination leaderboard, which asks each model to summarize over 7,700 articles using only the facts inside them. In its latest snapshot, GPT-5.4 and Gemini 2.5 Pro both sit at 7.0%, Claude Sonnet 4.6 at 10.6%, and Gemini 3 Pro Preview at 13.6%.

The high numbers come from a different kind of test. Stanford's AI Index responsible-AI chapter charts hallucination rates on the AA-Omniscience benchmark from 22% for the best model to 94% for the worst. That benchmark asks 6,000 tough questions with no document to lean on and counts how often a model answers wrong instead of declining.

The takeaway nobody puts first: the model you pick matters less than the task you give it. Hand it the source and AI hallucinations drop to single digits. Ask it to remember and you're rolling dice.

One more pattern from the leaderboards: newer and "smarter" doesn't always mean fewer hallucinations. Several reasoning models score worse than their older siblings on the summary test, because longer thinking gives them more room to add things the document never said.

Real AI hallucinations that cost money

Short answer: The best-documented damage is in courts, where judges have now addressed AI-invented citations in 2,184 decisions worldwide. Outside law, AI hallucinations have forced a consulting refund on a government report, made an airline honor a policy its chatbot invented, and created a new hacking trick built on fake software names.

Courts: my breakdown of 2,184 cases

The case that made headlines first involved a New York lawyer who filed a brief citing six court decisions ChatGPT had invented. Since then, legal researcher Damien Charlotin has tracked every ruling he can find where a judge dealt with hallucinated material. I filtered his AI Hallucination Cases database on October 11, 2026. Here's who filed the AI-invented material:

Three things about these AI hallucinations jumped out when I ran the numbers. The biggest group isn't lawyers but people without one, using chatbots as a free attorney. The United States accounts for 1,492 cases (68%), followed by Canada (233), Australia (112) and the UK (76). And 1,791 rulings involved a fully fabricated source, not a slightly wrong one.

The pace of AI hallucinations in court is what worries me most. A freeCodeCamp analysis counted 1,633 cases in the database by mid-June 2026. Reaching 2,184 by October 10 works out to about 34 new rulings a week, roughly five every day.

Business: refunds and liability

Deloitte Australia agreed to partially refund an A$440,000 government report after a university researcher found references to papers that don't exist and a fabricated quote from a court judgment, AP reported. The revised report disclosed that a generative AI tool was used in its preparation.

In Canada, a tribunal ordered Air Canada to pay a customer after its website chatbot invented a bereavement-refund rule. The airline argued the chatbot was responsible for its own words. The tribunal disagreed: a chatbot is part of the company's website, and the company owns what it says, as Yahoo News Canada reported.

Code: fake packages become real attacks

AI hallucinations in code have their own attack name: slopsquatting. In a study of 576,000 code samples from 16 models, 19.7% of recommended software packages didn't exist. Worse, 43% of those fake names came back every time the prompt was rerun, which lets attackers predict them, register them, and fill them with malware.

Pattern across every case: nobody checked the one thing that was checkable. A case name, a reference, a policy page, a package registry. Each check would have taken under a minute.

How do you spot an AI hallucination?

Short answer: You can't spot AI hallucinations by tone, so check the parts that can be checked: names, numbers, quotes, links and citations. Be most suspicious of precise details on niche topics, sources you can't open, and facts that change when you ask the same question again.

Hallucination risk checker

I built this from the patterns of AI hallucinations above. Answer four questions about the AI answer in front of you, and it tells you how hard to check it. Nothing you click is saved or sent anywhere.

How do you prevent AI hallucinations? 7 fixes

Short answer: You can't reduce AI hallucinations to zero, but you can make them rare and easy to catch. Give the model the source, let it say "I don't know," make it quote its evidence, and check the few details that are checkable. These seven fixes work in ChatGPT, Claude, Gemini and Copilot.

Here's the wrapper I'd paste in front of any research question where AI hallucinations would hurt:

If you want a tool built around fix 4, Perplexity shows its sources next to every answer by default, which makes the "open the link" habit faster. Claude and NotebookLM can also answer from documents you upload. For my everyday Claude setup, see how I use Claude AI every day.

Which AI hallucinates the least?

Short answer: There's no single winner, because each benchmark measures a different kind of AI hallucinations. On answering hard questions from memory while admitting uncertainty, Claude Opus 5.5 and GPT-6 Astra lead Artificial Analysis's index. On faithful summaries, smaller models often beat flagships. Pick by task, not brand.

Leaderboards checked October 11, 2026. AA-Omniscience index runs from −100 to 100; Vectara snapshot dated May 11, 2026. Scores change with every model release.

My practical read on AI hallucinations: for anything from memory, a model that declines when unsure beats a model that always answers. For summaries of your own documents, almost any current chatbot is good enough as long as you make it quote. If you're choosing between search-first assistants, my Perplexity vs ChatGPT comparison covers how each handles sources.

Final verdict: living with AI hallucinations

Short answer: AI hallucinations aren't going away, but they're predictable. They cluster where the model works from memory on rare facts, and they shrink when you hand it the source. Treat every chatbot answer as a strong draft, and verify the names, numbers and citations before they leave your screen.

Back to the hook: the same tools that put invented cases in front of judges can summarize a document with few errors. The difference was never the brand. It was whether someone gave the AI a source and checked what came back. If you're curious how heavy AI use affects your own thinking, the ChatGPT brain study is a good next read.

📖 Companion Read

Rebooting AI: Building Artificial Intelligence We Can Trust Gary Marcus & Ernest Davis · why fluent AI still gets facts wrong

Written before chatbots went mainstream, it explains why systems trained on patterns produce confident nonsense. Reading it now, the predictions land hard.

Get the book →