The Impossibility of Eliciting Latent Knowledge

In three linesTheoretical paper on the Eliciting Latent Knowledge (ELK) problem: how to train AI systems to honestly report beliefs about hidden variables. Authors formalize ELK using Causal Influence Diagrams and prove an impossibility theorem: no feedback-based training strategy depending only on agent behaviour can guarantee agent honesty, even with perfect feedback.

Summary generated by Claude — human-verified