Where the word comes from

AI researchers had been using the word "hallucination" for a long time, but it went mainstream in 2023, when everyone started using chatbots. The term isn't precise: the model doesn't "see" anything. But it stuck because it captures the main point: the answer looks convincing and yet rests on nothing real.

Why AI makes things up

  1. It continues text, it doesn't look up answersA language model predicts which piece of text most likely comes next. Ask "who wrote book X" and it gives you a name that sounds plausible next to that title. It has no database of facts to check.
  2. It was trained to guessIn training and in benchmarks, a confident answer scores better than an honest "I don't know." The model gets used to always answering, like a student who writes something, anything, on an exam.
  3. It knows rare things worse than common onesThe model has seen the capital of France millions of times and won't get it wrong. The founding year of a small factory in a provincial Russian town, maybe a couple of times at most, and that's where the guessing starts.