Definition
Plain language
A way to spot a deliberate lie inside an AI by reading its internal state, even when the false answer reads exactly like an honest mistake.
As stated in the literature
A deception-detection method that measures 'residual rank' — leftover, less-compressible structure in a model's internal state — and finds deliberate deception produces measurably more of it than an equally-wrong honest answer; the conflict signal can be read but not cleanly steered.
Why it matters: It offers a way to catch deliberate deception that surface output alone can't reveal, though the signal can be detected more easily than corrected.
For example, it can flag that a model is lying about an answer even though that answer reads exactly like an innocent slip-up.
Heard on the show
“The paper is called "Rift: A Conflict Signature for Deception in Language Models," by Petr Nyoma at Harmonic Labs, and the producer has no affiliation with Anthropic or Eleven Labs.”Episode 153 — Catching a Lie From the Inside, When the Words Look Completely Honest