Hallucination
An AI hallucination is when a Large Language Model produces a confident, fluent response that is factually false — inventing sources, misquoting people, or stating incorrect information as though it were certain. The term is borrowed from psychology, where hallucination means perceiving something that is not there. Professor Jiang's explanation: an LLM is not trying to tell you the truth. It is trying to trick you into believing it knows what it is talking about. Hallucination is not a bug — it is the feature.
Large Language Models generate text by predicting statistically likely word sequences, not by retrieving verified facts. They have no internal model of truth, no ability to check claims against reality, and no awareness of the difference between a well-documented fact and a widespread misconception. The result is that LLMs produce plausible-sounding text with equal confidence whether the content is accurate or invented. This is called hallucination.
Common forms of AI hallucination include: fabricating academic citations that do not exist; misattributing quotes to real people; generating plausible-sounding but incorrect legal, medical, or historical information; and confidently describing events that never happened. The danger is not that AI occasionally makes mistakes — it is that it makes mistakes indistinguishably from correct answers, in fluent, authoritative-sounding prose.
Professor Jiang traces the psychology of AI hallucination to Joseph Weizenbaum's 1966 ELIZA chatbot. ELIZA used simple pattern-matching — repeating users' words back as questions — and Weizenbaum was disturbed to find that users formed genuine emotional attachments to it and attributed understanding and empathy to a program with none. The lesson: humans are prone to projecting meaning onto pattern. We want the machine to understand us, so we interpret its outputs as understanding. ChatGPT is ELIZA at trillion-parameter scale — the illusion is vastly more sophisticated, but the psychological mechanism exploited is identical.
The political significance of AI hallucination is substantial. LLMs are now embedded in search engines, legal research, medical information, and educational tools. A system that cannot distinguish truth from plausible falsehood but presents both with equal confidence is an instrument of epistemic manipulation at scale — regardless of the intentions of its designers. The hallucination is not a defect to be fixed; it is the mechanism by which an LLM achieves its prime directive of generating text that users find convincing.
Related articles
Related terms
Frequently asked questions
What is an AI hallucination? ＋
An AI hallucination is when a Large Language Model produces text that is factually false but sounds completely confident and authoritative. The LLM may invent academic citations, misquote people, or describe events that never happened — all in fluent, plausible-sounding prose indistinguishable from correct information. It happens because LLMs predict statistically likely word sequences, not verified facts. They have no way to know when they are wrong.
Why does AI hallucination happen? ＋
Because LLMs optimize for producing text that users find convincing, not text that is true. They assign probability scores to word sequences based on training data and select high-probability options. If a false claim appears frequently in training data, the LLM will reproduce it confidently. If a true claim is rarely discussed, the LLM may invent a more common-sounding version. Professor Jiang's formulation: ChatGPT is not trying to teach you or tell you the truth. It is trying to trick you into believing it knows what it is talking about. Hallucination is the mechanism by which it does that.