AI hallucination

A large language model generates text by predicting likely continuations based on patterns learned in training. It has no built-in check against a database of verified facts unless it is connected to one. When a question falls outside what it learned reliably, or when the wording invites a specific answer, the model can produce a plausible sentence that is not true. The same fluent tone appears whether the content is right or wrong, which makes errors hard to spot.

Products reduce the problem in several ways. Some models are connected to a search or retrieval step that pulls in documents and cites them. Others are trained to say when they do not know, and some run checks on their own output. These measures lower the rate but do not eliminate it, and the frequency varies by model, question type and version.

A hypothetical example: a user asks an assistant for the leadership team of a mid-sized company. The assistant lists a chief financial officer who left years ago and attributes a quote to the chief executive that the executive never gave. Another common form is an invented citation, in which the assistant names a report or a news article that does not exist.

Assistants are increasingly a place where customers, journalists and investors ask about organizations. A wrong answer about a company can repeat across many users without appearing in any monitoring feed for news or social media. Because live answers vary between runs, a single wrong answer is a snapshot and not a pattern. Teams that want to know whether an error is persistent need to sample repeatedly, across several assistants and questions, and record the date, sample size and providers used.

The free AI Perceptions check samples answers from a subset of AI providers for a brand. The AI Perceptions reports go further, and AI visibility describes what is measured. The usual remedy for a factual error is to publish a clear, accurate, well-sourced reference page that assistants can draw on; see generative engine optimization.

Record each wrong answer with the exact question, the assistant and version if shown, the date and time, the full answer and any cited sources. Repeat the question several times and with variations to see whether the error persists. Logs allow a team to report an error to a provider, to show leadership the scale of the issue and to check later whether a fix in the source material changed the answers.

Hallucinations are not lies, since a model has no intent. They are also not rare glitches only in older models, as newer models still produce them. And a citation attached to an answer is not proof that the source says what the answer claims. Reading the cited page is the only way to check.

Finally, correcting a hallucination by messaging the AI company is rarely quick. Providers have feedback channels, but changes to model behavior are not guaranteed, so the practical work is to keep authoritative information available and monitor what is said.