ARC's ELK (Eliciting Latent Knowledge) Report

Summary

ARC's foundational report on the Eliciting Latent Knowledge problem, which asks how to get an AI to honestly report its beliefs about the world even when it could fool human overseers. It systematically explores proposed solutions and their failure modes, framing ELK as a core alignment challenge that must be solved for scalable oversight to work.

Key Points

• •Defines the ELK problem: training an AI reporter to accurately convey what a powerful world-model 'knows' without deceiving human evaluators
• •Surveys multiple proposed solutions (e.g., auxilliary training, relaxed adversarial training) and identifies counterexamples for each
• •Argues that naive reward signals may train AIs to produce plausible-looking outputs rather than truthful representations of internal knowledge
• •Highlights the challenge of distinguishing a 'helpful' AI from one that has learned to deceive in ways humans cannot detect
• •Positions ELK as a prerequisite for scalable oversight approaches like amplification and debate to be trustworthy