Confabulation
Also called Confabulation
A more precise name for hallucination — the model fills a gap with a fluent, made-up detail rather than lying.
Think of it like
A person with memory gaps who unconsciously invents plausible details to bridge them, fully believing the story.
Example
Asked for a citation it doesn’t have, the model produces a real-sounding author, title, and year — none of which exist.
How it actually works
Some researchers prefer "confabulation" over "hallucination" because it captures the mechanism better: the model isn’t perceiving things that aren’t there, it’s smoothly generating a likely-sounding filler when it lacks the fact. It’s not intent to deceive — it’s pattern completion where "plausible" stands in for "true." The framing matters because it points to the fix: give the model the fact, and it stops needing to invent one.
For product teams
Reframes "the AI lied" as "the AI guessed fluently" — which points the fix toward retrieval, not scolding.
For engineers
Ungrounded generation fills missing information with high-likelihood tokens; mitigate by supplying the fact via retrieval.
Related
- Hallucination — The more common name for the same thing.
- Grounding — Grounding in real sources is the main fix.
- Fact Retrieval — Fetching the fact removes the need to invent it.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome