LLMs Don't Hallucinate -- They Make Mistakes

In this video, I argue that large language models (LLMs) don’t “hallucinate” in any literal sense. The term began in computer vision to describe useful gap-filling, but it has drifted into LLMs, where it now refers to factual errors. I trace the history of the word, give examples, and explain why the metaphor is misleading. Hallucinations are perceptual experiences without external stimuli—something only minds can have. LLMs don’t perceive the world; they generate text. Calling their mistakes “hallucinations” hides what’s really happening, feeds hype, and discourages clarity.

I consider objections, show why the analogy fails, and consider alternatives like “error,” “bug,” or “confabulation.” If we want transparency in AI, we need to stop using the wrong metaphors and call these mistakes what they are.