Hallucination
When a language model generates fluent output that is false or unsupported by its sources, presented as if it were fact.
Hallucination is why the agentic web values grounding, structured data and verifiable sources: an agent that can retrieve a fact does not have to invent one.
- term
- Hallucination
- category
- core
- short_def
- When a language model generates fluent output that is false or unsupported by its sources, presented as if it were fact.
- long_def
- A hallucination is a confident but ungrounded generation: the model predicts plausible tokens rather than retrieving a verified fact, so the answer can be wholly invented — a citation, a number, an API that does not exist. It is a property of next-token prediction, not a bug, which is why grounding and retrieval are used to constrain it.
- see_also
groundingragai-agent- etymology_origin
- — verify-against-primary-at-build ↗ https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence) — term borrowed from human perception; adopted for generative models in the late 2010s, no single coiner
- related_to
groundingragagentic-ragai-agent- contrast_with
- Unlike a factual error copied from a bad source, a hallucination is generated with no source at all — the model fabricates it from statistical likelihood.
- example
- An agent asked for a library's API might invent a method that reads plausibly but does not exist; grounding the agent in the real docs is what prevents it.
- source
- https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence)
- status
- active
- why_it_matters
- Hallucination is why the agentic web values grounding, structured data and verifiable sources: an agent that can retrieve a fact does not have to invent one.
- sameAs
https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence)- bridge_entity
- grounding
- last_verified
- 2026-07-06
- md_twin
- /glossary/hallucination.md
last verified · by Özden Erdinc