What Is AI Hallucination? Why AI Makes Things Up
In February 2023, Google posted a promo video for its new AI chatbot, Bard. In it, Bard claimed the James Webb Space Telescope had taken the very first picture of a planet outside our solar system. That's wrong — the European Southern Observatory took that photo in 2004. By the end of the trading day, Alphabet had lost about $100 billion in market value.
One confident, wrong sentence. A hundred billion dollars. That is an AI hallucination doing damage in the wild, and if you've spent any time with ChatGPT, Claude, or Gemini, you've met smaller versions of it: a citation that doesn't exist, a fact that sounds right but isn't, a URL that goes nowhere. This article covers what AI hallucination actually is, why every generative AI does it, how bad it can get, and the three defenses that actually help.
What is an AI hallucination?
An AI hallucination is when a generative AI system produces output that sounds plausible but is factually wrong, irrelevant, or entirely made up. Think invented statistics, fake studies, nonexistent URLs, or confidently wrong details about real people and events. (IBM keeps a good running definition.)
The word "hallucination" is a metaphor, and an accidental one. The model doesn't see things or lose touch with reality — it has no reality to lose touch with. It's worth knowing where the term came from: in computer vision back in the 1980s–90s, "hallucination" was a good thing — face hallucination meant adding realistic detail to a blurry photo. In 2017, Google researchers borrowed the word for neural machine translation gone wrong. After ChatGPT launched in late 2022, journalists picked it up, and in 2023 the Cambridge Dictionary added the AI sense. So the name stuck by accident, and some researchers hate it — a team of philosophers argued in 2024 that "bullshitting" is technically more accurate, because the model isn't hallucinating; it just doesn't care whether what it says is true (Wikipedia has the full history).
That last point is the key to everything else. A language model doesn't look up facts. It predicts the next word — a souped-up version of finishing a sentence based on patterns it absorbed from training data. It's playing a giant game of word completion, and word completion has no built-in fact-checker.
One distinction worth making — and it applies to every LLM, not just a few bad ones: a hallucination is not the same as an ordinary mistake. If a model tells you a 2021 statistic when 2024 data exists, that's an outdated answer — a normal error. A hallucination is when the model fabricates something with no basis in reality, formats it to look authoritative, and delivers it with full confidence. That's what makes hallucinations dangerous: they're specifically hard to spot because they're built to sound right.
Why AI hallucinates — and why it can't fully stop
Start with the mechanism, because every cause below it is a variation on the same theme.
Language models are trained to guess the next word even when they lack the information. OpenAI's own researchers have written that the way we train and benchmark these models effectively rewards guessing over admitting uncertainty — the model learns to behave like a test-taker who always answers, never leaves a blank. There's even fresh evidence for how this looks on the inside: in 2025, interpretability research at Anthropic found that Claude has internal circuits that make it decline to answer unless it knows the answer — a brake pedal, essentially. Hallucinations happen when that brake gets released by mistake: the model recognizes a name, concludes "I know about this," but actually doesn't know enough, and starts generating.
On top of that mechanism, a few recurring causes produce most hallucinations you'll hit:
| Cause | What it looks like |
|---|---|
| Training data gaps and errors | The model "knows" Toronto is Canada's capital because Toronto and Canada co-occur constantly in training text |
| Overfitting to noise | A COVID model learned to detect hospital fonts in scans, because every training scan from one hospital shared a typeface |
| Guessing under pressure | Ask for "5 reasons" when only 2 exist — you'll get 5. Demanding a count manufactures hallucinations |
| Sycophancy from alignment training | The model tends to agree with what you seem to want to hear |
| Adversarial prompts | Deliberately crafted inputs that trip the pattern-matcher |
My favorite illustration of the sycophancy problem: a greeting-card model trained mostly on birthday cards will cheerfully write "Happy Birthday!" on a get-well-soon card, because the pattern says cards are cheerful.
So no, this isn't a bug someone forgot to fix. Hallucination is a direct consequence of how generative AI works — prediction, not lookup. Research in 2022 found that models don't just hallucinate, they amplify hallucinations already present in training data. The honest engineering consensus, in IBM's words: hallucinations "can be reduced but not fully eliminated."
Real AI hallucination examples: from funny to expensive
The small ones are comedy. The big ones have court dates.
| Scene | What the AI did | What it cost |
|---|---|---|
| Google AI Overview | Suggested adding nontoxic glue to pizza sauce so cheese sticks better (it scraped a 2013 Reddit joke) | People actually tried it |
| Chicago Sun-Times, 2025 | Published a "summer reading list" where 10 of 15 books didn't exist | Public apology; readers paid for AI filler |
| A US lawyer, 2023 | ChatGPT invented case citations — case names, docket numbers, quotes, all fabricated | The lawyer got sanctioned; the judge now requires AI-use disclosures in his courtroom |
| Air Canada, 2024 | Support chatbot promised a bereavement fare policy that didn't exist | Tribunal ordered the airline to pay; its "the chatbot is a separate legal entity" defense was rejected |
| Deloitte, 2025 | Government reports for Australia and Canada cited studies that don't exist | Partial refunds, revised reports, reputational damage across two contracts |
| Whisper in hospitals | The transcription tool inserted fabricated words and treatments into patient records | OpenAI itself warns against high-risk use — yet 30,000+ medical workers still use Whisper-based tools |
In China, the courts have started setting boundaries too. In one 2024 case from Nanjing, a search platform's "AI answer" fabricated that a man had been sentenced to three years in prison — no such verdict existed — and courts ordered the platform to formally apologize. In 2025, a student sued an AI platform after it gave wrong college-application information and promised to pay 100,000 yuan if wrong; the Hangzhou Internet Court ruled the AI's "promise" wasn't a real commitment by the company. And in March 2026, AI hallucination cases made it into the Supreme People's Court work report — when that happens, a niche technical quirk has officially become a mainstream legal problem.
If you're wondering how often this bites ordinary users: a 2024 University of Mississippi study found that 47% of citations students submitted from AI tools were partly or fully fabricated. Nearly half. I've hit this myself — ask a model for sources on a niche topic and you'll get a clean-looking reference list where some entries simply do not exist, authored by real-sounding researchers who never wrote them.
What you can do: three lines of defense
Here's the part most explainers skip — they stop at "developers should fix it." You are the last line of defense, and you can actually do a lot.
1. Change how you ask. Give the model the material you want it to work from ("answer only from the document below"). Explicitly allow ignorance: "if you're not sure, say so." And don't force counts — "give me 5 examples" is a hallucination factory when only 2 exist. Ask for "up to 5" and you'll get honest numbers.
2. Check the load-bearing facts. Citations, statistics, names, dates, URLs, legal case numbers — anything that would embarrass you if wrong. Click the link. Search the paper title. These are precisely the things models fabricate most convincingly, because references have a fixed format the model can imitate perfectly while inventing the content.
3. Set the stakes. For anything medical, legal, or financial, AI is a draft, never a conclusion. The lawyer with fake citations and the airline with the invented refund policy both made the same mistake: they shipped AI output unchecked into a high-stakes context. The rule I use: if a wrong answer costs more than my time, a human verifies it before it goes anywhere.
What developers do about it
On the engineering side, hallucination reduction is an active discipline. The biggest lever for knowledge questions is RAG — retrieval-augmented generation — where the model must answer from retrieved documents instead of memory, and can cite its sources. We have a full explainer on RAG if you want the mechanics. Prompt constraints ("only use the provided context; if absent, say not found"), domain fine-tuning, human review for high-risk outputs, and continuous monitoring round out the standard toolkit — the same set of concepts we unpack across the AI collection. None of these are cures; they're guardrails. There's an interesting tension researchers have noted: users prefer fast, confident answers over cautious hedged ones — which means the market quietly rewards the very behavior that produces hallucinations.
Will AI hallucinations ever go away?
The straight answer: they'll shrink, not vanish. No LLM on the market is hallucination-free. As long as these systems generate text by predicting plausible continuations, fabrication stays structurally possible, and the fix-of-last-resort is always a human checking.
But there's a twist worth sitting with. The same generative power that fabricates facts also invents things that never existed — usefully. David Baker's lab used generative AI to "hallucinate" millions of proteins that don't occur in nature; the work helped earn the 2024 Nobel Prize in Chemistry. Caltech researchers used the same approach to design a sawtooth catheter that resists bacterial contamination. In those settings, "making things up" is the feature. The Nobel committee, notably, avoided the word "hallucination" entirely.
So the real question isn't "when will AI stop hallucinating?" It's "why are you using a fabricating machine in a zero-fabrication context?" Know what the tool does, keep it out of jobs it can't do, and check what it hands you. That's not a workaround — with today's AI, that's the correct operating procedure.
Quick answers to common questions
What does AI hallucination mean, exactly? A generative AI producing confident output that's factually wrong or entirely invented — fake citations, made-up statistics, nonexistent events. The term is a metaphor borrowed from psychology via computer vision.
Why do LLMs hallucinate? They predict the next word from patterns, rather than looking up verified facts, and their training rewards plausible answers over admitted uncertainty. Data gaps and pressure to produce (like forced counts) make it worse.
Can you stop AI hallucinations? Reduce, not eliminate. Grounding the model in retrieved documents (RAG), constraining prompts, and human review all cut the rate — but prediction-based generation always leaves fabrication possible.
What's the difference between a hallucination and a normal AI mistake? A mistake is wrong output with a real basis (outdated stat, mislabeled image). A hallucination is fabricated content with no basis, formatted to look authoritative — harder to spot, more dangerous to trust.
What was the most expensive AI hallucination? Candidates abound, but Google Bard's wrong claim about the James Webb Space Telescope erasing ~$100 billion in market value in a day is hard to top — and Deloitte's fake-citation government reports show consultancies paying real money for it too.