What is AI model hallucination and how to reduce it?

A hallucination is an answer that sounds confident and plausible but is false or not supported by any source: an invented quote, a study that does not exist, a wrong date or price. The model is not lying on purpose. It produces the most likely continuation of the text, and when it does not know, it often guesses.

In a paper published in September 2025, OpenAI researchers argued that training and evaluation reward guessing over admitting uncertainty, much like an exam that gives no points for "I don't know".

What reduces hallucinations in practice:

• explicitly allow the model to say it does not know;
• give it the source text and ask it to answer only from that text;
• with long documents, ask for word-for-word quotes first and the analysis second;
• ask for a source for every claim and drop claims that have none;
• check names, numbers and dates yourself before you publish anything.

Web search and RAG help, but they move the problem to the quality of the sources rather than solving it.