A hallucination is a statement from an AI model that sounds convincing but is false or made up. The model cites a study that does not exist, a number that appears nowhere, or a quote nobody ever said, all in the same confident tone as a correct answer.
Picture a student in an oral exam who cannot answer a question. Instead of staying silent, they say something that sounds good and roughly fits the topic. Ask a model without web search for “three studies on coffee and concentration, with links” and the same thing can happen: real-looking titles, real-looking authors, links that go nowhere.
Below, you will find out why language models hallucinate, how often it happens in independent tests, which kinds exist and what actually helps.
1. Why AI models hallucinate
A language model is not a database. At every step it calculates which token is most likely to come next. Likely does not mean true. Three causes work together.
- Gaps in knowledge. On rare topics, people or events, the model has seen little. Anything that happened after its knowledge cutoff it does not know at all. The prediction keeps running anyway and fills the gap with something that fits.
- Guessing is rewarded. In 2025, OpenAI researchers argued that training and benchmarks reward guessing. A guessed answer is sometimes right and scores, while “I don't know” never does. Models learn to give an answer rather than admit uncertainty.
- Randomness in the choice. The model does not always pick the single most likely token. It picks at random among several candidates, and parameters such as temperature control how much. Once a wrong detail has been drawn, the model builds on it consistently.
On top of that, a model does not reliably know what it knows. If you ask whether it is sure, the answer is only partly reliable, because it is predicted text too.
2. How often models hallucinate
There is no single hallucination rate, because it depends heavily on the task. A good impression comes from AA-Omniscience, a knowledge test by Artificial Analysis. It asks thousands of difficult knowledge questions and counts how often a model gives a wrong answer instead of admitting it does not know. The hallucination rate is the share of wrong answers among all answers that were not correct.
Our benchmark database holds values for 60 models, using the setting with the lowest hallucination rate for each model. The median rate is 40%. Only 5 models stay below 20%, and 24 sit above 50%. So they guess in more than half of the cases where they do not know the answer.
The range runs from 4% to 94%. A model that knows a lot is not automatically one that hallucinates little. What matters is whether it says “I don't know” when it does not know. How individual models perform is shown in the profiles in the model directory.
One caveat:
The test measures hard knowledge questions without any tools. When a model summarizes a text that sits in the prompt, the error rate is usually far lower. The numbers show the risk when you let a model answer from memory, not its error rate on every task.
3. Types of hallucinations
- What happens
- The model names studies, books, links or rulings that do not exist
- Example
- A study title with real-sounding authors and a link that leads nowhere
- What happens
- A number, date or name is incorrect
- Example
- A wrong founding year or an invented revenue figure
- What happens
- A summary or translation adds something that is not in the original
- Example
- A summary mentions a deadline that is missing from the contract
- What happens
- Software, code libraries or commands are described with abilities they do not have
- Example
- A code sample calls a function that does not exist in the library
- What happens
- Correct information is assigned to the wrong person, company or version
- Example
- A quote is credited to the wrong politician
gradually.aiResearchers often distinguish two groups. Intrinsic hallucinations contradict the material the model was given. Extrinsic hallucinations add something that the material neither supports nor refutes.
4. Worked example: same task, two prompts
A constructed example shows how much the prompt affects the risk. The outputs are illustrations and do not come from a specific model run.
Prompt A:
Which rulings exist on liability for AI-generated text in the United States?
Give case numbers.
Possible answer:
Key is a ruling by the U.S. District Court for the Southern District of New York
of April 31, 2024 (Case No. 1:24-cv-04127) ...Court and case number look real. In this example only the date gives the error away, because there is no April 31. Real hallucinations usually come without such a tell. Details like these are dangerous because they create credibility. This became widely known in 2023, when a New York lawyer cited cases in a suit against the airline Avianca (Mata v. Avianca) that ChatGPT had invented.
Prompt B:
Here are three rulings as PDFs. Summarize what they say about liability for
AI-generated text. Cite the page for every statement.
If the documents do not answer a question, write "not included".Prompt B hands the model the material, demands evidence and explicitly allows it to admit a gap. That lowers the risk considerably but does not rule out errors. For important statements, check the cited passage anyway.
5. What helps against hallucinations and what does not
- Helps most, because the model does not have to answer from memory
- Helps, because you can check claims one by one
- Helps, because the model guesses less often
- Helps with logic and math, but does not remove invented facts
- Catches some errors, but can be wrong too
- Barely helps on its own when the model lacks the knowledge
- Only partly reliable, the self-assessment is predicted text as well
gradually.aiRAG is no cure-all either. In a 2024 Stanford study, legal research tools hallucinated in 17 to 33% of queries despite using RAG. If the search retrieves the wrong passages, the model simply builds its answer on the wrong material. And fine-tuning on new facts can even encourage hallucinations, as a 2024 study shows.
6. What this means for you
- Give the model the material. Upload the file, paste the text or switch on web search instead of asking from memory. How much fits in at once is covered in the entry on the context window.
- Always check numbers, quotes, names and sources. These are the places where hallucinations show up most often and do the most damage.
- Open every link. A source you cannot find may not exist.
- Allow gaps. Tell the model in your prompt or system prompt to admit missing information.
- Be extra careful with new and rare topics. The less there is about a subject online, the higher the risk.
- Keep the final decision. AI delivers a draft, and you decide what gets published or sent.