Hallucination
Also: 幻覺 · AI 編造 · 一本正經地胡說八道 · confabulation
The model produces text that reads perfectly plausibly but is factually wrong or invented.
When you will meet it
This is the most expensive risk in deploying AI. It is not "sometimes wrong" but "indistinguishable when wrong" — the tone is identical to when it is right.
An analogy
Like a supremely confident colleague who never says "I don't know". He does not stammer when wrong, so you can only catch it by checking, never by listening to his tone.
Minimal example
問：請列出三篇討論 X 的論文
答：1. Smith et al. (2021), Journal of X, 12(3), 45-67 ← 期刊、卷期、頁碼俱全
實際上這篇論文不存在。格式完整度與真實性無關。The more detailed an answer looks, the more it needs checking. Plausible detail is hallucination's most common disguise.
What people get wrong
- Trying to fix hallucination by switching models. It is a property of the training paradigm and architecture, not one vendor's bug.
- Treating "ask it for sources" as a safeguard. It will invent the sources too. The safeguard is that YOU check them, not that it claims them.