Hallucination
Noun · AI & Machine Learning · Origin: 2018
Definitions
Hallucination in artificial intelligence refers to a model generating content that is factually incorrect, fabricated, or nonsensical while presenting it with the same confidence as accurate information. This is a fundamental challenge with large language models (LLMs) because they generate text by predicting statistically likely next tokens rather than retrieving verified facts from a knowledge base. Hallucinations can range from subtle inaccuracies, like incorrect dates or misattributed quotes, to entirely invented citations, people, or events. Mitigation strategies include retrieval-augmented generation (RAG), which grounds responses in external documents, chain-of-thought reasoning that makes the model show its work, and fine-tuning with human feedback (RLHF). The term draws an analogy to human perceptual hallucinations. Reducing hallucination rates is one of the most active areas of AI research, especially for applications in healthcare, law, and finance where accuracy is critical.
In plain English: When an AI makes up facts or information that sound completely real but are totally wrong — and it says them with total confidence, making them hard to catch.
Example: "ChatGPT cited three court cases in a legal brief. All three were hallucinated — they never existed."
Retrieval-Augmented Generation (RAG) is the primary engineering approach to reducing hallucinations: ground the model's responses in retrieved documents rather than relying solely on parametric knowledge. But RAG doesn't eliminate hallucinations — it just makes them less frequent and easier to verify.
Example: 'Our legal AI hallucinates case citations that don't exist. We added RAG to ground responses in our actual case database, which reduced hallucination rate from 15% to 2%.'
Source: RAG mitigation
Etymology
- 2018
- NLP researchers begin using 'hallucination' to describe neural networks generating plausible but factually incorrect text
- 2023
- ChatGPT makes AI hallucinations a public concern; a lawyer submits hallucinated case citations to a federal court
- 2024
- Retrieval-Augmented Generation (RAG) and other techniques reduce but cannot eliminate hallucinations, making it the defining AI safety challenge
Origin Story
When AI confidently makes things up
The term 'hallucination' in AI predates the large language model era. It was used in computer vision research as early as the 2010s to describe when models 'saw' objects that weren't in an image. But the term exploded into mainstream use with ChatGPT's launch in late 2022.
As millions of people interacted with large language models for the first time, they discovered that these systems could generate perfectly grammatical, utterly confident, and completely fabricated information — citing papers that don't exist, inventing historical events, and making up legal precedents. Researchers debated the term: some argued 'hallucination' was too anthropomorphic, suggesting the model was perceiving something, when really it was just pattern-matching without grounding in truth.
Despite the criticism, 'hallucination' stuck because it viscerally captures the experience: the AI seems to be seeing something that isn't there. It became the defining concept of the generative AI era — the asterisk on every impressive demo.
Context: Computer vision research, 2010s; popularized 2022-2023
Fun fact: A 2023 legal filing became infamous when a lawyer submitted a brief containing case citations that ChatGPT had hallucinated — none of the cited cases existed.