Hallucination is fluent, confident language that is unsupported by reality, incorrect, or nonsensical. Large language models can sound intelligent while mainly recognizing patterns in internet text rather than reasoning about the physical world. section 10.1 section 10.2.3
The problem persists even as models grow larger: scaling improves grammar, vocabulary, and memorization, but does not reliably produce common-sense reasoning. Because text-only models lack a sense of reality and opportunities to learn from physical mistakes, fluent output can remain disconnected from the world. section 10.1 section 10.2.3
Counting cow legs with GPT-2
Given “There are 2 cows and 2 bulls, how many legs are there?”, ten generated answers included “four,” “2,” “30,” “three,” and “no.” The variation shows how an ungrounded model can produce plausible-looking language without consistently solving a simple real-world question. section 10.1
Use hallucination as a warning about language-model outputs:
- Treat confidence and fluency as separate from correctness. section 10.1
- Expect common-sense failures when a model has only consumed and produced text. section 10.2.3
- Ground outputs in information retrieval or real-world sensing when reliability matters. section 10.1 section 10.2.3
