Hallucination Detection
Also called Hallucination Detection
Automatically flagging when a model’s answer is likely made up rather than grounded.
Think of it like
A lie detector for the model’s own output — not to punish it, but to catch the confident guesses before they reach the user.
Example
Before showing an answer, a system checks each claim against the retrieved sources and flags any sentence the documents don’t actually support.
How it actually works
Since hallucination never fully goes away, the practical move is to catch it. Approaches include checking claims against source documents, asking the model to cite and verifying the citations, sampling multiple answers and looking for disagreement, or using a second model as a judge. None is perfect — the same fluency that makes hallucinations convincing to people can fool automated checks too — but detection plus a fallback beats blind trust.
For product teams
The safety valve that lets you ship a model that sometimes makes things up without shipping the mistakes.
For engineers
Flag ungrounded claims via source-consistency checks, citation verification, self-consistency sampling, or judge models; imperfect but pairs with fallbacks.
Related
- Hallucination — The failure it catches.
- Grounding Check — A specific check against sources.
- Output Validation — Part of validating outputs.
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