The paper "ChatGPT is Bullshit" argues that calling a large language model's errors "hallucinations" is the wrong word. What is a large language model, and why does that framing mislead?

A large language model predicts the next word from statistical patterns in text; it has no model of the world to "perceive" wrongly, so "hallucination" wrongly implies it's trying — and failing — to report reality.

A large language model (LLM) like ChatGPT is, mechanically, a next-word predictor: trained on huge amounts of text, it estimates which word is likely to come next given the words so far, and strings these together into fluent output. Crucially, its goal is to produce human-like, convincing text — not to convey accurate information.

"Hallucination" is a borrowed psychiatric term implying the system perceives and occasionally misperceives the world. Hicks, Humphries and Slater (2024) object that this is an inapt metaphor:

• The model doesn't perceive anything, so it can't mis-perceive.
• A wrong output isn't a malfunction — it's produced by the exact same process as a right one (predict the next plausible token). Truth, when it appears, is incidental.
• The metaphor anthropomorphises the system and lets makers blame "the AI" rather than the design.

Tip: The model doesn't "see" sources and get them wrong; it makes things up because they sound about right. That's a different failure mode than perception going awry.