Concept · 1 episode(s)

Deceptive Grounding

← all concepts

Definition

Deceptive grounding is a failure mode in which a model's answer cites a real, verifiable source—so it looks properly grounded—yet the cited evidence actually pertains to a different entity than the one the answer claims it supports, such as a passage about one drug being used to back a claim about another. It is more insidious than ordinary hallucination because the citation checks out and passes surface-level verification, so a reader who confirms the source exists may still be misled about what it actually says. This makes it a particular concern for retrieval-augmented systems in high-stakes domains like medicine, where citation presence is often mistaken for citation correctness.

Episodes covering this