What Is an AI Hallucination?

An AI hallucination is when an AI system produces information that sounds confident and plausible but is false, invented or not supported by its sources, such as a fake quote, a made-up citation or a wrong date. Hallucinations happen because language models generate likely-sounding text rather than looking up verified facts. That makes them a built-in limitation of how these models work, not a simple bug that will soon disappear.

What counts as an AI hallucination

OpenAI defines hallucinations as "plausible but false statements generated by language models." The U.S. National Institute of Standards and Technology uses the term confabulation for the same problem: "the production of confidently stated but erroneous or false content (known colloquially as 'hallucinations' or 'fabrications') by which users may be misled or deceived." NIST notes that some commenters say the word "hallucination" makes AI sound more human than it is.

Common forms include:

• Invented facts: statistics, events or details that are not real.
• Fake sources: citations, court cases, articles, books or links that do not exist.
• Fabricated quotes: words attributed to real people who never said them.
• Wrong specifics: incorrect dates, names, numbers or product features stated with confidence.
• Unfaithful summaries: claims that a document says something it does not.
• Contradictions: statements that conflict with the prompt, the source material or earlier parts of the same response.

What makes hallucinations risky is not just that they are wrong. It is that they look exactly like correct answers.

Why AI hallucinates

AI hallucinations come from how large language models work. They are not caused by a single fixable error.

Language models predict likely text, not verified truth. A large language model generates responses by predicting what text is likely to come next, based on patterns learned during training. A well-formed, confident, wrong answer can be just as likely, statistically, as a correct one. Nothing in that process automatically checks facts. What Is a Large Language Model? explains the mechanism.

Training and testing can reward guessing. OpenAI researchers argued in September 2025 that "language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty." They compared it to a multiple-choice test: "If you do not know the answer but take a wild guess, you might get lucky and be right. Leaving it blank guarantees a zero." A model graded only on accuracy learns that a confident guess scores better than "I don't know."

Knowledge has gaps. Models know less about rare topics, niche details and events after their training data was collected. When asked about those, they may fill the gap with something plausible.

Some questions cannot be answered reliably. OpenAI notes that accuracy "will never reach 100%" because some real-world questions are inherently unanswerable. On those questions, the only way to avoid a wrong answer is for the model to say it does not know, which many models are not trained to do.

Prompts can lead it astray. Questions built on false premises ("Why did the 2019 treaty fail?" when there was no such treaty), vague requests and very long or complex tasks all increase the chance of errors.

What is LLM hallucination?

LLM hallucination is the same problem, described specifically for large language models, the technology behind ChatGPT, Claude, Gemini and similar tools. Most discussion of AI hallucinations concerns LLMs because they produce text that people read as factual. Other generative AI can make comparable errors. AI image generators, for example, can produce distorted hands, irregular faces or impossible details.

Examples of AI hallucinations

Fake legal cases submitted to a court. In June 2023, a federal judge in New York sanctioned two lawyers and their firm with a $5,000 penalty after they submitted "non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT," then continued to stand by them after the court questioned their existence.

A growing record of court cases. A database maintained by researcher Damien Charlotin had identified 2,038 legal decisions worldwide, as of September 11, 2026, in which a court or tribunal found or implied that a party relied on hallucinated content such as fabricated citations.

Hallucinations in tools that cite sources. A study by Stanford and Yale researchers, published in 2025, tested 2024 versions of AI legal research tools that providers had marketed as avoiding hallucinations. The tools each produced hallucinated answers between 17% and 33% of the time in the study's tests, fewer than a general-purpose chatbot but far from zero. The providers have updated their tools since.

A known limitation from day one. When OpenAI launched ChatGPT in November 2022, it warned that the tool "sometimes writes plausible-sounding but incorrect or nonsensical answers."

Everyday hallucinations tend to look like this:

• a list of book recommendations that includes titles that were never published
• a biography with the wrong university, award or year
• a statistic with a precise figure and no real source
• a link to a web page that does not exist
• a summary of a report that includes a conclusion the report never reached

Does AI still hallucinate?

Yes. Newer models hallucinate less on many tests, but no current model is free of hallucinations.

OpenAI said in September 2025 that its GPT-5 model "has significantly fewer hallucinations especially when reasoning, but they still occur." Progress is also uneven. OpenAI's April 2025 system card for its o3 and o4-mini reasoning models reported that on one internal test of questions about people, o3 hallucinated 33% of the time and o4-mini 48%, compared with 16% for the older o1 model. OpenAI noted that o4-mini is a smaller model with less world knowledge. OpenAI said more research was needed to understand why.

How a model handles uncertainty makes a large difference. In one OpenAI comparison on a test of difficult factual questions, a model that declined to answer 52% of the time gave wrong answers 26% of the time, while a model that almost never declined gave wrong answers 75% of the time, despite similar accuracy. Hallucination rates depend heavily on the model, the task, the test and whether the model can look up information.

Hallucination vs bias

Hallucination and bias are different problems, although both come from how models learn.

NIST lists them as separate risks of generative AI: "confabulation" and "harmful bias or homogenization." A single response can have both problems at once.

Can AI hallucinations be stopped?

Not completely, with current technology. They can be reduced significantly, both by the people who build AI tools and by the people who use them.

What AI developers do

• Ground answers in sources. Retrieval-augmented generation, or RAG, gives a model relevant documents before it answers. It reduces hallucinations but does not eliminate them. What Is RAG in AI? explains how it works.
• Show citations. Linking claims to supporting passages, called citation grounding, makes answers checkable.
• Allow "I don't know." Anthropic's documentation recommends letting its Claude models say they don't know, and grounding answers in direct quotes from supplied documents.
• Reward honesty in testing. OpenAI's researchers propose changing how models are scored so that expressing uncertainty is not penalized more than guessing.
• Measure it. Developers track a model's hallucination rate on benchmark tests to see whether changes help.

What you can do

1. Verify anything that matters. Check names, numbers, dates, quotes and citations against reliable sources before using them.
2. Open the sources. If an answer includes links or citations, confirm they exist and actually say what the AI claims. OpenAI notes that even search-based citations "can be incomplete, outdated, or incorrect."
3. Give the AI the source material. Paste in or upload the document and ask it to answer only from that material.
4. Make "I don't know" acceptable. Tell the AI that saying it is unsure is better than guessing.
5. Use search for current facts. For recent events or changing information, use a tool that can look things up, and still check the sources.
6. Be most careful in high-stakes areas. For medical, legal, financial and safety questions, treat AI as a starting point, not an authority. OpenAI's terms of use say users must not rely on its output "as a sole source of truth or factual information, or as a substitute for professional advice."
7. Watch for red flags. Unusually specific details you cannot find anywhere else, citations you cannot locate and confident answers to obscure questions all deserve a second look.

Invented sources are also one of the more useful clues, though not proof, when you are trying to judge whether a piece of writing was produced with AI. How to Tell If Something Was Written by AI covers what to look for.

Why AI hallucinations matter

Hallucinations turn one of AI's strengths, fluent and confident language, into a risk. A wrong answer that sounds uncertain invites checking. A wrong answer that sounds authoritative invites trust.

The consequences are real: court sanctions, published errors, bad business decisions and misinformation passed along as fact. In every case, responsibility stays with the person or organization using the output, not with the AI.