Fact-checking in the AI context is the practice, and the tooling, of verifying that model outputs are supported by trusted sources before they reach users.
AI fact-checking verifies claims against evidence: extract checkable assertions, retrieve authoritative sources, compare claim to evidence (entailment/support scoring), and route low-confidence cases to humans. In generation pipelines it runs as a verification layer: checking model output against its cited sources before display.
Generation scales faster than verification, so automated checking is the only way trust keeps up: both for content ecosystems flooded with synthetic text and for enterprises whose assistants must not invent policy. 'Generate, then verify' is the 2026 architecture for high-stakes output.
As a filter, yes: verification against provided sources catches a large share of unsupported claims and prioritizes human review. As an oracle, no: checkers share failure modes with generators, so high-stakes pipelines keep humans on the low-confidence tail.
Groundedness asks 'does the output match its given sources?': tractable and standard in RAG. Open fact-checking asks 'is this true in the world?': harder, needing source discovery and credibility judgment.
Between generation and exposure: verify claims against sources, attach confidence, block or flag failures, and log everything for audit. Stakes determine strictness: from silent logging to mandatory human sign-off.