In plain words
Ask a chatbot for three books on a topic and it may give you two real ones and one that was never written, complete with a believable title and author. That invented answer is called a hallucination.
It happens because a language model does not look facts up. It writes the most likely-sounding next words, based on patterns from its training. Most of the time likely and true line up. When the model doesn't really know, it can still produce something that sounds just right.
The tricky part is the tone. A hallucination sounds exactly as sure as a correct answer, so you cannot spot it by how it reads. Names, numbers, dates, quotes and references deserve the most checking.
Everyday example
The student who never leaves a blank
In an exam where a blank answer gets zero, a student who doesn't know a date writes a plausible one anyway: "1857, I think" becomes simply "1857". Sometimes it is right. Over many exams, guessing earns more marks than honesty. AI models were trained and scored in a similar way, so they learned to guess confidently too.
Where the example breaks: The student knows they are guessing. The model has no separate feeling of knowing or not knowing. It just produces the likeliest text, which is why it cannot reliably warn you.
How it actually works
OpenAI's September 2025 paper defines hallucinations as "plausible but false statements generated by language models" and traces them to how models are trained and tested.
Training rewards fluency, not truth
In pre-training, a model learns to predict the next token in huge amounts of text. Nothing marks each sentence as true or false. Regular patterns like spelling get learned well. Rare one-off facts, such as an obscure person's birthday, cannot be inferred from patterns, so the model fills them in.
Tests reward guessing
Most benchmarks score only accuracy. A guess sometimes scores, "I don't know" never does. OpenAI's paper argues this pushes models to "guess over acknowledging uncertainty", and suggests scoring that penalises confident errors more than honest abstentions.
Grounding reduces it
Giving the model real sources to read, through web search or retrieval-augmented generation, makes answers far more anchored. It still can misread or misquote a source, so citations need a click.
What you can do
Ask for sources and open them. Ask the model to say when it is unsure. Paste in the document you want answered from. Example: "Answer only from the text below. If the answer isn't there, say so."
- Definition
- "Plausible but false statements generated by language models" (OpenAI, 2025)
- Famous case
- In 2023 a US federal court in New York fined lawyers $5,000 after they filed a brief citing court cases ChatGPT had invented (Mata v. Avianca)
- Warned from day one
- ChatGPT's November 2022 launch post said it "sometimes writes plausible-sounding but incorrect or nonsensical answers"
- Other name
- Some researchers prefer "confabulation", a term from psychology for filling memory gaps with invented details
Hallucination is different from bias (skewed patterns from data) and from sycophancy (agreeing with you to please you), though all three produce wrong answers.
Questions people ask
Why does ChatGPT make things up?
Because it writes the most likely-sounding text rather than looking facts up. When its training didn't cover something well, likely-sounding and true can split apart. Training and tests have also rewarded confident guesses over saying "I don't know".
Can AI hallucinations be fixed completely?
Not completely, as of October 2026. Newer models hallucinate less, and search and retrieval help a lot. But any system that generates text from patterns can still produce a confident error, so checking important claims remains necessary.
How can you tell if an AI is hallucinating?
You usually cannot tell from the wording. Check names, numbers, quotes and references against a real source. Be most careful with obscure topics, recent events, exact figures and citations, where invented details are most common.
Is a hallucination the same as the AI lying?
No. Lying means knowingly saying something false. The model has no intent and no inner check of truth. It produces plausible text, and sometimes plausible text is wrong. The effect on you can be the same, which is why checking matters.