A grounding check confirms an answer is backed by real evidence, not invented.
DefinitionWhat it means
A grounding check is a verification step, run by a rule, a classifier, or another model, that traces each factual claim in an output back to a trusted source document or database record before the answer is shown to a user.
Why it mattersWhy you should care
Grounding checks are the last line of defense against hallucination in high-stakes AI products such as legal, medical, or financial assistants, and enterprises increasingly require visible citations plus an automated grounding check as a condition of deploying a generative feature at all.
At a glanceSee it
Opens the verify step — the answer is split into atomic claims, each scored against retrieved evidence past a support threshold, so a claim counts as grounded only when the evidence actually entails it.
The control layer around the verdict — a fully backed answer ships, an unsupported one loops back to regenerate against evidence, and repeated failure ends in a safe abstain rather than a guess.
Where you see itIn the wild
- Citation features in enterprise search and document Q&A products.
- Compliance reviews for regulated industries requiring traceable answers.
- Eval discussions distinguishing grounding checks from faithfulness scoring.