An AI hallucination is when an artificial intelligence system — usually a chatbot like ChatGPT — produces an answer that is confident, fluent, and completely false. The model doesn't stammer or flag its uncertainty. It states the fabrication with the same authority it uses for the truth, which is exactly what makes hallucinations so dangerous. A made-up statistic, a fake quote, a court case that never existed — all delivered in the same smooth, believable voice.
Hallucinations aren't a rare glitch you can patch away. They're a built-in side effect of how today's language models work. Understanding why they happen is the key to using AI tools safely — and to knowing when you absolutely must double-check what the machine tells you. This guide explains what hallucinations are, why they occur, what they cost in the real world, and the proven techniques that reduce them.
What Is an AI Hallucination?
A hallucination is any output where an AI model presents false or fabricated information as if it were fact. The word is borrowed from human psychology, though the comparison is loose — the model isn't "seeing things," it's generating plausible text that happens to be wrong.
Crucially, a hallucination is different from a simple mistake. If you ask a model a genuinely ambiguous question and it gets it wrong, that's an error. A hallucination is more specific: the model invents something — a source, a date, a name, an event — that has no basis in reality, and presents it with total confidence. It fills a gap in its knowledge with a convincing guess rather than admitting it doesn't know.
This behaviour is a direct consequence of what these systems are: pattern-based text predictors, not databases of verified facts. To see why that matters, it helps to understand the machinery underneath — which we cover in what machine learning is and, more specifically, in how ChatGPT actually works.
Why Do AI Models Hallucinate?
The single most important thing to understand about a large language model is that it does not look facts up. It predicts the next most likely word, over and over, based on patterns it absorbed during training. That's it. There is no fact-checking step, no internal library it consults, no sense of "I'm not sure about this one."
Compare it to a search engine. When you search, the engine retrieves real documents that exist and shows them to you — the answer is traceable to a source. A language model does something fundamentally different: it generates text that sounds like a correct answer, based on statistical patterns. Most of the time those patterns line up with reality, so the answer is right. But when the training data is thin, outdated, or contradictory, the model still produces a fluent, confident response — it just fills the hole with an invention.
Several factors make hallucinations more likely:
- Knowledge gaps. Ask about something rare, recent, or niche, and the model has fewer real patterns to draw on — so it improvises.
- A bias toward being helpful. Models are trained to give answers. Saying "I don't know" is often under-rewarded, so they'd rather guess than refuse.
- Biased or flawed training data. A model learns from whatever it was fed. If that data contains errors or bias, the model reproduces — and sometimes amplifies — them.
- Vague prompts. An unclear or leading question gives the model room to wander and invent.
A Real Example: The Lawyer Who Trusted ChatGPT
The risks aren't hypothetical. In 2023, a New York attorney named Steven Schwartz used ChatGPT to research a personal-injury case, Mata v. Avianca. The chatbot produced a polished legal brief citing at least six court decisions that supported his argument perfectly.
There was one problem: none of the cases were real. ChatGPT had fabricated them — complete with fake case names, fake quotes, and fake internal citations to other invented rulings. When opposing counsel and the judge tried to locate the cases, they couldn't, because they didn't exist. When the court asked Schwartz to produce the decisions, the model even generated fake full texts of them on request.
Judge P. Kevin Castel sanctioned the lawyers involved, ordering a $5,000 penalty and requiring them to notify the real judges falsely named in the fabricated opinions. The episode became a landmark warning about AI in professional work — not because the technology is useless, but because it was trusted without verification. The lawyer assumed a confident answer was a correct one. That is the exact trap hallucinations set.
Common Types of AI Hallucinations
Hallucinations come in a few recognisable flavours. Knowing them helps you spot one before it costs you:
| Type | What it looks like |
|---|---|
| Fabricated facts | Invented statistics, dates, or events stated as true |
| Fake sources | Citations, studies, or URLs that don't exist (as in the court case above) |
| Made-up quotes | Words attributed to a real person who never said them |
| False attribution | A real fact credited to the wrong source, company, or author |
| Contradiction | An answer that conflicts with itself or with the prompt |
The common thread is confidence. In every case, the model presents the fabrication in the same authoritative tone it uses for accurate information — which is why you can't rely on how sure it sounds.
How to Reduce AI Hallucinations
You can't switch hallucinations off entirely, but you can cut them down dramatically with a few habits and techniques. The most powerful is grounding the model in real information.
- Ground it in real sources (RAG). The most effective fix is retrieval-augmented generation, where the model is given authoritative documents to answer from — so it quotes real facts instead of guessing. We explain the technique in detail in our guide to retrieval-augmented generation (RAG).
- Ask for sources and verify them. Require citations or links you can check — then actually check them. A fabricated source falls apart the moment you try to open it.
- Lower the "creativity" for factual work. Many tools let you reduce randomness (often called temperature). A lower setting keeps the model on safer, more likely answers.
- Write clear, specific prompts. Give context, and explicitly tell the model to say "I don't know" when it isn't sure. Vague questions invite invented answers.
- Keep a human in the loop. For anything that matters — legal, medical, financial — a person must verify the output before it's used. The Mata lawyers skipped this step, and it cost them.
Can AI Hallucinations Be Eliminated Completely?
Not entirely — and this is important to accept. Research suggests hallucination is an inherent limitation of how language models work, not just a bug that better engineering will someday remove. Because a model generates plausible text rather than retrieving verified truth, there will always be some gap between "sounds right" and "is right."
The realistic goal, then, isn't a hallucination-free AI. It's a workflow reliable enough that fabrications rarely reach the final output — achieved by combining grounded retrieval, source verification, careful prompting, and human review. Model makers keep improving the raw accuracy, but the responsibility to verify hasn't gone anywhere. Treat AI as a fast, fallible assistant, not an oracle.
The Bottom Line
AI hallucinations are confident, fabricated answers that language models produce because they predict plausible text rather than looking up facts. They aren't rare accidents — they're a structural feature of the technology, and they can't be fully eliminated. But they can be managed: ground the model in real sources with RAG, demand and check citations, prompt carefully, and always keep a human verifying anything that matters. The lesson from the courtroom is simple and permanent — an AI answer that sounds certain is not the same as an answer that's true. For the bigger picture of how these systems work and where they fit, explore our Artificial Intelligence & LLMs hub.
Frequently Asked Questions
What is an AI hallucination in simple terms?
An AI hallucination is when an AI system, such as a chatbot, gives an answer that is confidently stated but false or made up. Instead of admitting it doesn't know something, the model fills the gap with a plausible-sounding invention — a fake statistic, a quote no one said, or a source that doesn't exist. The term describes any output where AI presents fabricated information as if it were verified fact, delivered in the same fluent, authoritative tone it uses for correct answers.
Why do AI chatbots make things up?
Because a language model predicts the next most likely word based on patterns in its training data — it does not look facts up or fact-check itself. When the data is missing, outdated, or unclear, the model still generates a smooth, confident response, filling the gap with a guess. It's also trained to be helpful and give answers, so it tends to invent something rather than say "I don't know." The result sounds convincing but can be entirely wrong.
Can AI hallucinations be completely eliminated?
No. Research indicates that hallucination is an inherent limitation of how large language models work, not merely a bug to be patched. Because these systems generate plausible text rather than retrieving verified facts, some risk always remains. The practical goal is to reduce hallucinations to a rare event using retrieval-augmented generation, source verification, careful prompting, and human review — not to expect a perfectly reliable AI. The duty to verify important answers stays with the human user.
What is a real example of an AI hallucination?
In 2023, a New York lawyer used ChatGPT to research the case Mata v. Avianca, and the chatbot produced a legal brief citing at least six court decisions — all fabricated, with fake names, quotes, and citations. Nobody could find the cases because they never existed. Judge P. Kevin Castel sanctioned the attorneys $5,000 and required them to notify the real judges falsely named. It became a landmark warning about trusting AI output without verification.
How can I reduce hallucinations when using AI?
Ground the model in real information: use tools that rely on retrieval-augmented generation (RAG), which answer from authoritative documents rather than memory. Ask for sources and actually verify them, since a fake citation collapses the moment you check it. For factual tasks, lower the model's randomness setting, and write clear prompts that invite an honest "I don't know." Most importantly, keep a human in the loop to verify anything consequential before you rely on it.
Are AI hallucinations the same as lying?
No. Lying requires knowing the truth and choosing to misrepresent it. An AI model has no concept of truth, intent, or deception — it simply predicts likely text, and a hallucination is what happens when the most likely-sounding text is false. There's no motive behind it. That distinction matters because it explains why hallucinations can't be fixed with "honesty" and why they appear in the same confident tone as accurate answers: the model genuinely can't tell the difference.
Sources
- What are AI hallucinations? — IBM
- ChatGPT cited bogus cases in a New York federal court filing — NBC News
- Mata v. Avianca, Inc. — Wikipedia
- Hallucination is Inevitable: An Innate Limitation of Large Language Models — arXiv
- A Comprehensive Survey of Hallucination Mitigation Techniques in LLMs — arXiv
