Faithfulness
Whether the model's output is faithful to (consistent with and supported by) the input context.
Think of it like
Like a book adaptation being faithful to the source material.
Example
Given a document, the model summarizes it. Faithfulness checks: did the summary avoid adding information not in the document?
How it actually works
Faithfulness is especially important for summarization and question-answering from documents. A model that adds plausible-sounding but unsupported details violates faithfulness. You measure it by checking whether each claim in the output is supported by the input. Harder for longer outputs (more claims to check).
For product teams
For RAG and retrieval-augmented tasks, faithfulness is critical. A model that mixes context with hallucinations is worse than one that admits uncertainty.
For engineers
Use LLM-as-judge or human raters to check faithfulness. Add retrieval-based guardrails to ground outputs in context.
Related
- Factuality — Related concept.
- RAG — What improves this.
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