Watch: Hallucination

When a large language model hallucinates, it produces text that sounds completely confident and fluent, but is actually false or entirely unsupported by any source.

This happens because these models generate the most probable next word based on statistical patterns in their training. They are not pulling from a database of verified facts, so a wrong answer reads just as convincingly as a right one. Whenever a model lacks a fact, has outdated information, or is pushed too far, it defaults to making things up. This can look like invented statistics, fake web links, or quotes confidently assigned to the wrong brand.

To fight this, developers use techniques like grounding and retrieval-augmented generation, also known as RAG. This process gives the model real sources to read and cite. But even with RAG, hallucinations can still happen if the model misinterprets the source or ignores it.

For brands, this cuts both ways. A false claim can quickly spread through AI answers. At the same time, if a brand is missing from the sources a model can retrieve, the AI is much more likely to guess and hallucinate details rather than sharing anything accurate.