When AI Makes Things Up Why chatbots hallucinate, why they sound so certain, and how to catch them The confident nonsense machine Ask an AI chatbot a question and it may deliver a beautifully structured answer, complete with dates, technical jargon and the unshakable confidence of a man explaining cryptocurrency at a dinner party. Occasionally, the answer is entirely wrong. This is an AI hallucination: a claim that is false, fabricated or unsupported, presented as though it belongs in the factual universe. What counts as a hallucination? Hallucinations range from invented scientific studies and imaginary quotations to wrong dates, fictional legal cases and citations that lead nowhere. More subtle examples stitch together genuine facts into a conclusion the evidence does not support. A mistaken inference is not necessarily a hallucination; the trouble starts when an unsupported inference is sold as established fact. Why does it happen? Large language models learn statistical patterns from vast amounts of text and generate responses token by token. That training gives them impressive language and reasoning abilities, but it does not equip them with a perfect internal fact-checker. When information is missing, ambiguous or obscure, a plausible continuation can win over an accurate admission of uncertainty. Training and evaluation that reward answering rather than abstaining can make this worse. The machine is not necessarily attempting deception; it is producing an output that can be wrong without reliably recognizing the error. Five ways reality goes missing First, pattern completion can manufacture an answer to a question with no answer, such as naming the winner of a future Nobel Prize. Second, incomplete information can encourage false connections between true facts. Third, ambiguous names and prompts invite the model to guess the intended subject. Fourth, long reasoning chains can compound a small early mistake. Fifth, requests for obscure numbers, research citations or up-to-the-minute developments can exceed what the model can reliably establish. So how often does ChatGPT hallucinate? There is no honest universal percentage. Error rates depend on the particular model, task, benchmark, question difficulty, access to tools and even how 'hallucination' is defined. Basic explanations of well-established concepts may be comparatively reliable; obscure quotations, highly specific statistics and recent events without verification are riskier. Mathematical and coding answers can also be wrong, though calculators, execution and tests can expose many errors. Any neat percentage offered without a named model and test method deserves immediate suspicion.Can AI check its own homework? Sometimes. Web search, trusted databases, calculators and code execution let an AI compare its output with evidence. Retrieval can reduce unsupported claims, but it is not a magical truth serum: sources can be outdated, contradictory or misread, and a citation can be real while failing to support the sentence beside it. Verification works best when the model checks the precise claim, not merely whether a vaguely relevant page exists. Are newer systems getting better? Many newer systems are trained to express uncertainty, use tools and avoid fabricated references. Those improvements can reduce errors on particular evaluations, but progress is uneven. A model that performs well on one benchmark may still confidently invent an obscure detail elsewhere. Fluency should never be mistaken for a warranty. Humans have an awkwardly similar problem Human memory is reconstructive, not a flawless video archive. People misremember names, merge events and repeat stories with total conviction. AI shares the outward symptom of confident error, though the underlying mechanisms differ. At least a human can be embarrassed at a reunion when the alleged eyewitness turns out to have been in another country. The practical survival guide Treat confident prose as presentation, not proof. Ask for primary sources when the stakes matter. Check that citations exist and actually support the claim. Separate established evidence from interpretation and speculation. Use calculations, code tests or authoritative records where possible. For medical, legal and financial decisions, verify consequential claims independently. Most importantly, reward the words 'not enough evidence' rather than demanding a heroic answer to every question. The punchline AI hallucination is not proof that the technology is useless. It is proof that an eloquent answer and a correct answer are different products. The most trustworthy system is not the one that never hesitates; it is the one that knows when the evidence has run out. Unfortunately, confidence remains considerably cheaper to manufacture than truth. Editorial note: This article describes general behaviour of large language models, not a measured hallucination rate for any specific release.