Glossary term
Glossary term
Foundations
When a model generates plausible-sounding but factually incorrect or unsupported content.
The production of plausible-seeming but factually incorrect output by a generative AI model that purports to be making an assertion about the real world. For example, a generative AI model that claims that Barack Obama died in 1865 is hallucinating.
Air Canada's AI chatbot hallucinated a non-existent bereavement-fare policy in 2024 - a passenger sued and won damages, establishing AI hallucination as a legal liability risk for businesses.
Amazon Lex's enterprise Q&A bot was found to hallucinate product specifications by a consumer electronics retailer - the team mitigated it by switching to RAG-grounded generation, reducing hallucination rate from 12% to 0.3%.
Vectara's Hughes Hallucination Evaluation Model benchmarks 11 LLMs on factual consistency - showing GPT-4o hallucinates in 3% of RAG-grounded responses vs. 15% for smaller models.
Created for this library
A legal-tech vendor reduces hallucination by adding strict retrieval grounding so the assistant cites a clause for every claim.
A medical writing assistant flags hallucination risk and shows the source passage so reviewers can verify every generated sentence.
An enterprise chatbot logs hallucination examples from user feedback to prioritize which knowledge sources to add to retrieval.
Definition source: Google for Developers Machine Learning Glossary | Creative Commons Attribution 4.0 License