Mistakes and Invented Facts
An answer can contain a mistake. Point it out in your next message: the model will usually agree and produce a corrected version.
The correction holds for that conversation, and the answers that follow are built with it. In a new chat the same mistake can come back word for word.
Where hallucinations come from
A hallucination is an invented number, date, quotation, link, statute number, or expert name, presented exactly like a verified fact.
Hallucinations are among the problems model developers work on seriously, and companies are trying to reduce them. For new versions they measure the share of invented content and publish the results along with the release.
The model does not pull facts out of a reference book. It composes every answer from scratch, choosing a likely continuation one word at a time, and a correct answer is produced the same way as a wrong one.
The model can admit a gap in what it knows, and newer versions do it more often than older ones. When data is missing, it will still sooner fill the gap with plausible text than refuse to answer.
A confident wording says nothing about accuracy. The invented part of an answer looks exactly like the verified part, so check anything that matters separately.
The risk of invention is higher where there was little data: a rare case in your industry, a recent event, an exact number, a link, someone's name.
An invented citation
You ask the model to back up your point with a study. What comes back is a citation: author names, a journal title, a year, an issue number, a link.
Every part of a citation like that looks plausible, and the whole was never checked against any catalog. The link can lead to a "document not found" page, and the author may turn out to be someone who never wrote the paper.
There is a separate case, when a model's answer gets retold as an expert opinion. The chapter "An Answer From a Superintelligence" covers it.
In a 2025 study, researchers had OpenAI's GPT-4o (omni) write six literature reviews on mental health.
Of the 176 citations in them, 35 pointed to papers that do not exist. Of the rest, 64 contained an error, most often in the number by which the article is found in a catalog.
These figures describe one model and one way of setting the task. The researchers deliberately gave the model a simple prompt, the kind an ordinary user would write, with no added demands such as verifying the references. Each prompt was run once.
OpenAI had opened internet search to all ChatGPT users, free accounts included, in December 2024, six months before the study. Whether it was on during the study is not stated in the paper.
The published prompts do not mention search, and the links in the answers point only to catalog numbers, with no pages found.
Behavior changes with how the task is set
Say it outright: "if you do not have enough data, say so." That puts refusal among the acceptable answers.
Grounding in sources works better still. Turn on internet search and ask for links: when an answer is built on a document that was found, the share of invented content drops sharply.
Checking before you send
A second model works well for a recheck. Give it the same input and compare the answers: a discrepancy shows you where to look for the error.
You can also ask the same model to fact-check itself. Have it attach to every claim a piece of evidence you can verify yourself: a link to a source and a quotation from it. Open the link and confirm the quotation is there.