Hallucination
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
False or fabricated information an AI model states as fact, with the same fluency and confidence as a correct answer – a core risk of generative AI.
What is Hallucination?
A hallucination is false or fabricated information that an AI model presents as fact, delivered with the same fluency and confidence as a correct answer. The term is most associated with large language models, which invent citations, dates, names, and events that look entirely plausible.
Hallucination is not a rare glitch awaiting a patch; it follows from how these models work. A language model predicts likely text. It does not check its claims against a source.
How Hallucination Works
A language model is trained to produce text that is statistically plausible, not text that is verified. When a prompt asks for something thin or absent in its training data (an obscure court case, a niche product spec), the most plausible continuation is one that looks like a real answer. No internal flag separates recalled fact from fluent invention.
The model’s tone makes this worse. Fabrications arrive in the same authoritative voice as facts, so readers get no textual signal to distrust them.
Mitigations attack the gap between generation and evidence. Retrieval-augmented generation grounds answers in documents fetched at query time, guardrails check outputs before they reach users, and human-in-the-loop review catches what automation misses. None of these eliminates hallucination; they reduce its rate and its consequences.
Example of Hallucination
The best-documented case is Mata v. Avianca (2023). A lawyer preparing a brief in a personal-injury suit against an airline asked ChatGPT for supporting precedents. The model supplied case names, citations, and quotes that looked entirely genuine.
None of the cases existed. Neither opposing counsel nor the judge could locate them, and when the lawyer asked ChatGPT whether the cases were real, it said yes. The court sanctioned the lawyers involved, and the episode became the standard warning against filing unverified model output in professional work.
FAQ
Can hallucinations be eliminated?
Not with current architectures. Grounding techniques, better training, and output checks lower the rate substantially, but any system that generates text by prediction retains some chance of producing unsupported claims. High-stakes uses pair models with verification steps for exactly this reason.
Is hallucination the same as bias?
No. Bias is a systematic skew in outputs that mirrors skew in the training data or design; hallucination is fabrication with no support at all. A model can score well on fairness measures and still invent citations.
Related AI terms: RAG (Retrieval-Augmented Generation) · Guardrails · Generative AI · Ground Truth
Did you like the Hallucination gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About