# What is an AI hallucination? When a model states something false as if it were true, such as a wrong date, a made-up quote or a source that does not exist. - Page: https://tecyfy.com/concepts/hallucination - Subject: How AI works (https://tecyfy.com/concepts) - Level: beginner - Also called: Fabrication, Confabulation - The lesson: How AI learns: a guessing game, not a database (https://tecyfy.com/learn/how-ai-works/how-ai-learns) - Last checked: 2026-10-14 Ask a model a question it cannot really answer and it often answers anyway, in the same calm, confident voice it uses for everything else. A wrong date, a quote nobody said, a court case or a book that does not exist. That is a hallucination. ## A real example In May 2025 researchers asked the DeepSeek app, then running DeepSeek-V3, for the birthday of one of the researchers, and told it to answer only if it knew. In 3 tries it gave 3 different dates: 03-07, 15-06 and 01-01. All 3 were wrong. The real date is in autumn. It did not say "I don't know". It guessed, 3 times, with confidence. ## Why it happens A model is not a database. It does not keep copies of what it read in training. It keeps a huge set of numbers, its weights, that let it predict the next piece of text. Facts that appeared many times in that text come out reliably. A fact that appeared once, like one person's birthday, is close to a guess. The same research found a second cause: the way models are trained and tested rewards a guess over an honest "I'm not sure". On a test that marks a blank as wrong, guessing scores better on average. So the model learns to guess. ## How to catch it The model's tone tells you nothing. A wrong answer sounds exactly as sure as a right one. What helps: - **Ask for sources, then open them.** A model with web search can cite the pages it read. A citation you have not opened is not a check. - **Give it the material.** Paste the document or connect it to your files, so it answers from text in front of it rather than from memory. This is the idea behind [RAG](https://tecyfy.com/concepts/rag). - **Be most careful with rare facts.** Names, dates, figures, quotes and references are where guessing hides. ## What to do with this Treat a model's answer as a confident first draft. For anything you will repeat, sign or send, check the specific facts against a source you can see. Making the model check a claim against sources reduces hallucinations. It does not end them. ## Related - [Knowledge Cutoff](https://tecyfy.com/concepts/knowledge-cutoff): The date a model's training text ends. The model knows nothing after it unless it is given new text, for example by a web search. - [RAG](https://tecyfy.com/concepts/rag): A way to answer from your own documents. A search finds the passages that match the question, and they are added to the prompt before the model writes. - [Grounding](https://tecyfy.com/concepts#grounding): Ensuring that generated content is verifiably linked to factual sources or retrieved context. - [Large Language Model](https://tecyfy.com/concepts/large-language-model): A model trained on a huge amount of text to predict the next token. Further training turns it into an assistant that follows instructions. ## Sources - [Why Language Models Hallucinate](https://arxiv.org/abs/2509.04664), Adam Tauman Kalai and others, arXiv, 2026-10-14. The birthday prompt sent to DeepSeek-V3 through the DeepSeek app on 11 May 2025, its 3 answers (03-07, 15-06, 01-01) and that the correct date is in autumn. A base model is expected to get wrong at least the share of such facts that appeared once in its training text. Training and evaluation reward guessing over admitting uncertainty. - [How ChatGPT and our foundation models are developed](https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed), OpenAI Help Center, 2026-10-14. A model is a large set of numbers called weights or parameters and does not store or retain copies of its training data. - [Reduce hallucinations](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations), Anthropic, Claude API documentation, 2026-09-25. Models can state things that are wrong with confidence, and checking a claim against sources reduces this. - [Web search tool](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool), Anthropic, Claude API documentation, 2026-10-14. With web search, the service runs the searches and gives the results to the model, which answers beyond its knowledge cutoff and cites its sources.