QQuestion: Why does AI make things up?

At its core, the language model behind an AI chatbot isn't looking facts up when it answers you. It writes by picking, one piece at a time, words that are likely to come next. A likely-sounding sentence isn't always a true one, and when an AI confidently says something false, that's called a hallucination.

This has shown up outside the lab more than once. AutoGPT, released in March 2023, drew attention as an AI that just gets on with a task once you give it a goal, but it also showed the public hallucinations and endless loops¹. Grokipedia, an AI-written encyclopedia launched in October 2025, was criticized for missing sources and misquotations².

So researchers have looked for ways to cut down on made-up answers. The RAG paper from May 2020 had the AI look up relevant documents and read them before answering³. Chain-of-Verification, published in September 2023, has the AI write a draft, come up with its own checking questions, answer them separately, and then revise⁴.

Even with those methods, checking still matters. AstaBrief, a model for writing science reports released in October 2026, measured how accurate its citations were, not just its answers — because in a report, what matters most is not how smooth the sentences are but whether the sources it cites really exist⁵. So when an AI answer gives you a date, a number or a source, open that source and check it yourself.