Someone asks an AI chatbot for a quick fact. The answer comes back instantly, worded clearly and with total confidence. It sounds right.
Later, it turns out the fact was wrong. Not slightly off, but invented. A date that never happened, a book that does not exist, a quote nobody said.
This is one of the most confusing things about AI for people just getting started with these tools. The tool did not stumble or admit it was unsure. It stated something false as if it were plainly true.
There is a name for this. It is called an AI hallucination, and it is one of the most important things to understand before trusting anything an AI tool says.
This guide explains what an AI hallucination actually is, why it happens, the kinds that show up most often, and the simple habits that catch them before they cause a problem.
1. What does "AI hallucination" actually mean?
A short, plain-English definition
An AI hallucination is when an AI tool produces information that is false or made up, but presents it as if it were true.
It is not a glitch or an error message. The tool is working exactly as designed. It simply generated something that sounds correct but has no basis in reality.
The word covers invented facts, fake sources, made-up quotes, and confident answers to questions that have no real answer.
Why it is called a "hallucination" when nothing is really being seen
The word is borrowed from human experience, where a hallucination means seeing or hearing something that is not there.
AI is not seeing anything. The term is a shorthand that researchers and companies adopted to describe output that looks real but is not.
Some experts argue it is the wrong word, because it makes the tool sound confused or unwell. A more honest description is that the AI is confidently filling a gap with a plausible guess.
What a hallucination looks like in everyday use
Most hallucinations are not dramatic. They are small, believable details slipped into an otherwise reasonable answer.
Examples include:
- a statistic with a precise-looking number that was never published
- a book or film recommendation for something that does not exist
- a website link that leads nowhere
- a historical date that is simply wrong
- a confident summary of a law or rule that does not apply
The danger is exactly that they look ordinary. Nothing flashes to warn the reader that a line was invented. It is one of the reasons so many common beliefs about what AI can do turn out to be wrong.
2. Why does AI make things up?
AI predicts the next word, it does not look up facts
Most AI chatbots work by predicting the most likely next word in a sentence, based on patterns learned from huge amounts of text.
That is the key point. The tool is not searching a database of verified facts and reading back the answer. It is assembling a response that fits the pattern of a good answer.
Usually that produces something accurate, because accurate text was common in what it learned from. Sometimes the most natural-sounding continuation is simply not true. This is the same underlying process behind how generative AI creates any kind of content.
There is no built-in step that checks whether an answer is true
A person writing an important email might pause and check a fact before sending. Standard AI tools have no equivalent moment built in.
There is no internal fact-checker deciding whether each sentence is correct before it appears. The words are generated and shown, confident tone and all.
Some newer tools can now search the web while answering, which helps. Even then, they can misread a source or invent a detail the source never contained.
Confident wording is not the same as a correct answer
This is the part that catches people out. AI writes a wrong answer in the same calm, fluent style it uses for a right one.
There is no nervous hedging when it is guessing. A fabricated court case is described with the same authority as a real one.
Confidence is a feature of how the tool writes, not a signal that the information is true.
3. What kinds of hallucination happen most often?
1. Invented facts, dates, and figures
The most common hallucination is a plain factual detail that is simply wrong. A wrong founding year, a misremembered population figure, a statistic that sounds specific but was never real.
These are the hardest to catch, because a single wrong number can sit inside an otherwise accurate paragraph.
2. Fake sources, studies, and citations
AI is strikingly good at inventing sources. It will produce a study title, an author name, a journal, and a date that all look real but do not exist as a set.
Made-up references are now showing up in published research. A 2026 study in The Lancet audited 2.5 million biomedical papers and found the share containing at least one reference that does not exist in any major database rose from 1 in 2,828 papers in 2023 to 1 in 277 in early 2026 (The Lancet, reported by STAT News).
The study tracks the rise, not the cause. It cannot prove AI wrote each fake reference. The timing lines up with wider AI use, but that is a pattern rather than proof.
3. Made-up quotes and people
AI can attribute a quote to a real person who never said it, or invent a person entirely and describe them in detail.
Because the wording is smooth and the name is often real, these fabrications are easy to accept without a second thought.
4. Confident answers to questions with no real answer
Ask AI something that has no settled answer, or something based on a false premise, and it will often answer anyway rather than say it does not know.
Instead of pushing back, it produces a tidy, confident response built around the flawed question. Learning how to write a clearer request reduces this, though it does not remove it.
4. Where do hallucinations cause the most harm?
Medical and health questions
A wrong answer about a symptom, a dosage, or a condition can lead someone to act on information that is simply incorrect.
AI can be useful for explaining a medical term in plain language. It should never replace a doctor for diagnosis, treatment, or anything urgent.
Legal and financial questions
Laws and financial rules are detailed, change often, and depend on personal circumstances. This is fertile ground for confident, wrong answers.
An AI response can sound authoritative while missing the rule that actually applies, which is why important decisions still need a qualified professional.
Anything with names, numbers, or sources attached
Hallucinations cluster around specifics. Names, dates, figures, and citations are exactly where AI is most likely to fabricate.
If an answer hangs on a precise detail, that detail is the thing to check first. It also helps to have a realistic sense of what these tools are genuinely good at before leaning on them.
5. Real examples of AI hallucinations going wrong
Fabricated legal cases that reached a courtroom
In 2023, a New York lawyer used ChatGPT to research a legal filing. It cited six past cases as precedent. None of them existed.
When he asked ChatGPT to confirm the cases were real, it doubled down and said they could be found in databases such as Westlaw and LexisNexis. They could not. The judge fined the lawyer, a second lawyer who filed the document, and their law firm $5,000 between them (Mata v. Avianca, sanctions opinion via Justia).
The case became a warning across the legal profession, and the pattern has kept repeating in courts since.
Invented citations in research and medical writing
The fake-source problem is measurable. In a 2024 test published in the Journal of Medical Internet Research, researchers asked AI tools to produce references for medical systematic reviews. GPT-4 got 28.6% of its references wrong, with at least two of the title, first author, or publication year incorrect. GPT-3.5 got 39.6% wrong, and Google's Bard 91.4% (Journal of Medical Internet Research).
Getting the source right is the entire point of a citation, and these tools failed at it routinely. That is why an AI-supplied citation means nothing until it has been checked against the original.
Everyday examples most people can relate to
Hallucinations are not only a problem for lawyers and scientists. In 2022, an Air Canada chatbot told a customer he could claim the airline's reduced bereavement fare back after he had already flown. That was not the airline's actual policy. A tribunal later ruled the airline had to honour what its chatbot wrongly promised (Moffatt v. Air Canada, 2024 BCCRT 149).
Even the companies building AI have been caught out. In 2023, Google's Bard chatbot claimed in a promotional advert that the James Webb Space Telescope took the first image of a planet outside our solar system. That was wrong, the first such image was taken in 2004, and the error was spotted within hours (Reuters, via NPR).
6. How can hallucinations be spotted and reduced?
1. Treat every answer as a starting point, not a final answer
The single most useful habit is to treat AI output as a first draft rather than a finished, trustworthy result.
A starting point is genuinely useful. A final authority is not what these tools are.
2. Check names, numbers, and sources against a trusted place
Because hallucinations gather around specifics, the specifics are where to look. Take any name, figure, date, or source and confirm it somewhere independent.
A quick search of the original study, the official website, or a known reference will catch most fabrications in seconds. There are also practical fixes for cutting down wrong answers in the first place.
3. Ask the tool to show where the information came from
Asking for the source of a claim can help, especially with tools that search the web as they answer. Following up with a second question is often what exposes a weak answer.
One warning: do not treat the tool confirming its own answer as proof. In the courtroom case above, the AI confirmed its own invented cases were real. A check has to go to an outside source, not back to the same tool.
4. Be extra careful when the answer sounds very confident
Confidence is not a signal of accuracy. If anything, a smooth and certain answer to a tricky question deserves more scrutiny, not less.
The safest approach is to slow down exactly when an answer sounds most convincing.
7. What AI hallucination means in practice
Hallucination is not a rare fault that only affects careless users. It is a normal feature of how these tools work, and it can appear in any answer, on any topic, at any time.
That does not mean AI should be feared or avoided. It means AI works best as a helper, not the final authority.
Anything that carries real weight, a health decision, a legal point, a figure going into a document, or a source going into an article, needs checking against something outside the tool before it is trusted.
Final thoughts
An AI hallucination is simply the tool stating something false as if it were true. It happens because AI predicts likely wording rather than looking up verified facts, and because nothing inside a standard tool checks whether an answer is correct before showing it.
The reason it catches people out is the confidence. A wrong answer arrives in the same fluent, certain style as a right one, with no warning attached.
The fix is not complicated. Treat answers as a starting point, check the specifics against a trusted source, and be most careful when something sounds most convincing.
Used that way, AI stays genuinely useful without quietly leading anyone into a mistake.
Frequently asked questions
Why does AI make things up when it does not know an answer?
Most AI tools are built to always produce a response, not to admit uncertainty. When the tool does not have a real answer, it generates the most plausible-sounding text instead, which can mean inventing a detail rather than saying it does not know.
Does an AI hallucination mean the tool is broken?
No. A hallucination is the tool working as designed. It predicts likely wording rather than retrieving verified facts, so a confident but false answer is a normal outcome, not a sign of a fault.
Do newer or more advanced AI tools hallucinate less?
Often yes, especially tools that can search the web while answering. They still hallucinate, though. Even the latest tools can misread a source or invent a detail, so checking remains necessary regardless of how advanced the tool is.
Can AI hallucinations ever be fully switched off?
Not at the moment. Hallucination comes from the core way these tools generate text, so there is no setting that removes it entirely. The risk can be reduced with careful use, but not switched off.
How can someone tell if an AI answer is a hallucination?
The clearest method is to check the specifics. Look up any name, number, date, or source against an independent, trusted place. If a detail cannot be confirmed anywhere else, treat it as a likely hallucination.
Is an AI hallucination the same as AI simply being wrong?
They overlap but are not identical. Being wrong can mean giving an outdated or mistaken answer. A hallucination is more specific: the tool invents something that was never real and presents it confidently as fact. For more on why AI gets things wrong in general, see Can AI Be Wrong?.
