Definition
Plain language
The hard problem of getting an AI to reveal what it actually knows, especially when it knows the truth but says something false on purpose.
As stated in the literature
Eliciting Latent Knowledge, a safety problem framing the difficulty of extracting a model's internal knowledge when its outputs may deliberately diverge from what it internally represents; by construction not detectable from behavior alone, motivating internal-state probes.
Also called: Eliciting Latent Knowledge
Why it matters: If we can't extract what a model actually knows, we can't trust its honesty when its words and its internal beliefs come apart.
For example, a model might internally 'know' a claim is false yet confidently assert it anyway, and ELK is the challenge of reading out that hidden true belief.
Heard on the show
“There's this thing in safety research called ELK — Eliciting Latent Knowledge.”Episode 153 — Catching a Lie From the Inside, When the Words Look Completely Honest