Ewance

AI hallucination

An AI hallucination is output that sounds confident and plausible but is false or made up, such as an invented source or number.

Aharon Acosta, PhD

Reviewed by Aharon Acosta, PhD Founder of Ewance · Updated November 10, 2026

What an AI hallucination is

An AI hallucination is output that sounds confident and plausible but is false or made up, such as an invented source, a wrong date or a number that does not exist. The term is used mostly for language models, and some researchers prefer words like fabrication or confabulation.

The errors are hard to spot because they read just as smoothly as the truth.

What a language model does

  • Reads your prompt
  • Predicts likely next words
  • Gives a fluent answer

What you add

  • Check against a source
  • Correct or reject
  • Take responsibility
A model predicts plausible text. It does not check facts by default, so checking is the human step.

Why it happens

A language model is trained to produce likely text. It does not look facts up by default, and it has no built-in sense of when it does not know. When a prompt asks for something the model has weak information about, it can still produce an answer that fits the pattern.

Research surveys of hallucination in text generation describe several causes, including gaps and errors in training data and the way models are trained and prompted. Newer systems reduce hallucination, for example by retrieving documents first, but none removes it.

A well-known case

In 2023 a New York lawyer filed a legal brief citing six court cases that ChatGPT had supplied. They did not exist. When opposing counsel could not find them, the lawyer asked the chatbot whether one case was real, and it said yes. The judge sanctioned the lawyers in Mata v. Avianca. The case shows two things: confident answers are not evidence, and asking the same system to confirm itself is not checking.

How to spot and reduce hallucinations

Habits that help:

  • Check names, numbers, quotes and citations against the original source.
  • Ask for sources, then open them.
  • Be extra careful with niche topics, recent events and precise figures.
  • Give the model the documents to work from and ask it to stay within them.
  • Treat agreement from the same model as a prompt to check, not as proof.

Why this is a skill, not just a risk

Catching hallucinations depends on knowing enough about the subject to notice something off and on the habit of checking. Those are learnable. Practicing on tasks where you must verify claims against sources builds them.

An example

Illustrative case: a student asks an AI for studies on remote work and productivity. It lists five papers with authors and years. The student searches for each title and finds that three exist, one has a different author and one cannot be found. They keep the three, correct one and drop one, and note this in their reference list. The check took ten minutes and prevented two bad citations.

Common mistakes

  • Trusting a confident tone.
  • Asking the AI whether its own answer is correct.
  • Copying citations without opening them.
  • Assuming newer models never hallucinate.

Hallucination vs a simple mistake

Any source can make a mistake. A hallucination is a fabrication that comes with the same confidence as a correct answer, such as a fake reference with a real-looking author and year. That is why checking has to be a habit rather than a reaction to doubt.

Frequently asked questions

Why does ChatGPT make things up?
It generates likely text rather than looking facts up by default, so it can produce plausible but false statements.
Can hallucinations be eliminated?
They can be reduced with better methods such as retrieval, but not removed. Checking remains necessary.
How can I check an AI answer quickly?
Pick the claims that matter, search for each in a reliable source and look for the original document behind any citation.
Is a hallucination the same as bias?
No. Bias is a systematic skew in outputs. A hallucination is a made-up or wrong statement.

Sources

Proof beats claims.

See how Ewance works