Groundedness is the degree to which the claims in a large language model's output are actually supported by the source material it was given, such as retrieved documents, reference manuals, or the conversation context. It is the counterpart of hallucination, in which a model produces content absent from its sources, and is used nearly interchangeably with faithfulness.
It is a core metric for evaluating systems such as retrieval-augmented generation (RAG) that answer by consulting external documents. Evaluation is often automated by splitting a response into individual claims and having another model judge whether each claim is verifiable in the source material, either with an LLM-as-a-judge approach or a natural language inference model. Unlike factuality, which asks whether a statement is true in the world, groundedness considers only agreement with the provided material.
In psychology the word also describes a stable, centered state of mind.