Key takeaways
- OpenAI defines hallucinations as plausible but false statements, and NIST calls the same problem confabulation because the software is not conscious or lying on purpose.
- In OpenAI's SimpleQA results, o4-mini answered almost every question and was wrong 75% of the time, while gpt-5-thinking-mini declined 52% of questions and was wrong 26% of the time.
- In Mata v. Avianca, a federal judge fined two lawyers and their firm $5,000 after a brief cited court cases that ChatGPT had made up.
In this story 7 sections
Ask a chatbot for the title of a researcher's PhD dissertation and you might get a neat, specific answer that is completely wrong. Ask again and you could get a different wrong answer, delivered with the same calm confidence. That is an AI hallucination, and it is one of the most common ways people get burned by AI tools. Below are straight answers to the questions people search most about it, from what it means to how you can catch one before it costs you.
What does AI hallucination mean?
An AI hallucination is a false statement that an AI model presents as if it were true. OpenAI describes hallucinations as "plausible but false statements generated by language models," and that word plausible is the key. The answer reads well, fits the question and sounds like it came from someone who knows.
The National Institute of Standards and Technology uses a different word. Its Generative AI Profile calls the problem confabulation, the production of confidently stated but erroneous or false content. NIST notes that "hallucination" and "fabrication" imply the software is conscious or lying on purpose, which gives it too much credit. A chatbot is not seeing things and it is not trying to trick you. It is producing text that looks right without any built in sense of whether it is.
If you want the background on the technology itself, our plain English guide to what a large language model is covers how these systems are built.
Why does AI hallucinate?
AI hallucinates because it is trained to predict likely words, not to check facts. During pretraining, a model reads huge amounts of text and learns to guess the next word. That text carries no labels saying which sentences are true and which are false.
Patterns that repeat everywhere, like spelling or closing a parenthesis, get learned well. Rare, arbitrary facts do not. In its research post Why language models hallucinate, OpenAI points out that a fact like a pet's birthday cannot be predicted from patterns alone, so the model fills the gap with something that merely sounds right.
There is a second cause that gets less attention. OpenAI argues that the way models are graded rewards guessing. Most benchmarks score accuracy only, so a model that says "I don't know" gets zero, while a model that guesses sometimes gets lucky. OpenAI compares it to a multiple choice test where leaving a blank guarantees no points. Guess a stranger's birthday and you have roughly a 1 in 365 chance of being right. Decline to answer and your chance is zero. Over thousands of test questions, the guesser looks better on the scoreboard, even though it is wrong far more often.
Why does AI hallucinate sources, references and case law?
Citations are exactly the kind of detail that pattern prediction gets wrong. A model has seen millions of references and knows what one looks like: author names, a year, a title, a journal or a court. It can produce something with that exact shape without the source ever existing.
Legal research shows the risk clearly. In June 2023, Judge P. Kevin Castel of the U.S. District Court for the Southern District of New York sanctioned two lawyers and their firm in Mata v. Avianca. One of the lawyers had used ChatGPT for research, and the brief cited court opinions that did not exist, complete with fake quotes. When he asked the chatbot whether the cases were real, it said they were. The court imposed a $5,000 penalty and required letters to every judge whose name had been attached to a fake opinion.
NIST's profile also flags legal confabulations as a widespread problem in large language models. The lesson carries over to medical, academic and financial citations too.
What are common AI hallucination examples?
The most common hallucinations are invented specifics: dates, numbers, quotes, names and sources. Here is what they tend to look like in everyday use:
- A biography with the right career but the wrong birthday or hometown
- A quote attributed to a real person who never said it
- A study, book or article title that sounds real but cannot be found
- A statistic with a precise decimal and no traceable origin
- A software feature, menu path or setting that does not exist in your version
- A summary of a document that adds a point the document never made
That last one matters. NIST counts output that drifts from the source you supplied, or contradicts something the model said earlier in the same chat, as confabulation too. So even a summary of your own file can include made up material.
How often do AI chatbots hallucinate?
It varies a lot by model and by question, and the tradeoff is not always obvious. OpenAI shared results from a factual quiz called SimpleQA, taken from its GPT-5 system card:
- gpt-5-thinking-mini declined to answer 52% of questions, got 22% right and got 26% wrong.
- OpenAI o4-mini declined only 1% of the time, got 24% right and got 75% wrong.
The older model scored slightly higher on accuracy but was wrong roughly three times as often, because it almost never admitted it did not know. That is why a single accuracy number on a leaderboard tells you little about how much you can trust a chatbot. OpenAI says newer models hallucinate significantly less, but that they still do it.
Can you stop AI from hallucinating completely?
No, not completely, but it can be reduced a great deal. OpenAI lists several common myths in its research. Accuracy will never hit 100%, because some questions have no knowable answer. At the same time, hallucinations are not inevitable on every uncertain question, because a model can be trained to say it is unsure. And a bigger model is not required: OpenAI argues a small model can find it easier to know its limits.
Its proposed fix is to change how models are scored, penalizing confident errors more than honest uncertainty. Developers also connect chatbots to real documents so they answer from retrieved text instead of memory, a method explained in our guide to what RAG in AI is. Retrieval helps, though it does not remove the need to check.
How can you spot and avoid AI hallucinations?
Treat every specific fact as a claim to verify, and check the ones that matter before you act on them. A few habits catch most problems:
- Open the source. If a chatbot cites something, click through and confirm the page exists and actually says that. In Google Gemini, a Sources button appears under some responses and opens the related links in a side panel, according to Google's Gemini help page. No button means no links were provided for that answer.
- Ask the same question twice. If the details change between answers, as in OpenAI's dissertation example, the model is likely guessing.
- Give it permission to say "I don't know." Telling the chatbot to flag uncertainty, or to answer only from text you paste in, cuts down on confident guessing. Our tips on writing better AI prompts show how to phrase that.
- Be strict with high stakes topics. Legal, medical, tax and money questions deserve a primary source or a professional, not just a chatbot reply.
- Never ask the AI to confirm itself. The lawyers in Mata v. Avianca asked ChatGPT if its cases were real, and it said yes.
For a fuller routine, see our walkthrough on checking AI answers before you trust them, and browse more explainers in our AI section.
One quick test before you rely on any chatbot answer: pick the single most specific detail in it, a name, a number or a citation, and try to find it somewhere else in under a minute. If you cannot, assume it may be made up until you can.
Sources
Share this story
Spotted an error? Tell us


