what’s an LLM hallucination?

the term was borrowed from earlier work on neural machine translation and image captioning (where models would “see” objects that weren't there) and applied to LLMs as they began producing convincing-but-fabricated citations, statistics, and biographical details.

Ji et al.'s 2023 survey formalized two categories. intrinsic hallucinations contradict the input the model was given, like an LLM summarizing a document and inventing a fact that wasn't there. extrinsic hallucinations can't be verified against the input at all: pure fabrication that sounds plausible.

hallucinations matter for AI detection because they're often the most reliable tell in long-form AI-generated text: fabricated citations, invented court cases, plausible-but-wrong dates. a 2024 Stanford study found that general-purpose LLMs hallucinate on 58-82% of legal queries; even purpose-built legal-research tools hallucinate 17-33% of the time.

LLMs are trained to maximize next-token plausibility, not truth.

hallucinations are most dangerous when they sound right. confidence and fluency are produced by the same machinery as the rest of the text. a fabricated citation reads exactly like a real one.

retrieval-augmented generation (RAG) reduces but does not eliminate hallucination. even with cited sources in the prompt, models sometimes confabulate or misattribute. “the model said it cited this paper, but the citation doesn't exist” is a real failure mode in production RAG systems.

hallucination rate isn't a single number. it depends heavily on the task (factual recall, summarization, reasoning), the domain (legal/medical are worst), and how strictly you grade.

frontier models hallucinate less on common queries but can hallucinate more on niche or adversarial ones, and sounding confident does not mean being correct. as of 2026, hallucination remains built into next-token prediction across mainstream production LLMs.