Glossary · Term

deceptive grounding

← all terms

Definition

Plain language

When an AI answer is built entirely from real, correctly-cited sources but is still wrong because it attached that evidence to the wrong subject.

As stated in the literature

A retrieval failure where a model faithfully relays specific details from a genuine document but misattributes them to a different entity than the one queried, passing hallucination, faithfulness, and citation checks simultaneously.

Why it matters: It matters because an answer can pass every standard accuracy and citation check and still be confidently wrong, which is especially hard to catch.

For example, an AI answers a question about one senator by quoting real, accurately-cited voting records that actually belong to a different senator with a similar name.

Heard on the show

“The authors give it a name: deceptive grounding.”
Episode 214 — The Medical AI Answer That's Accurate, Sourced, and Still Wrong

Mentioned in 1 episode

  1. 214
    The Medical AI Answer That's Accurate, Sourced, and Still Wrong

Related terms