Glossary · Term

Korbak

← all terms

Definition

Plain language

Researchers who argued that reading an AI's step-by-step reasoning is a fragile but genuine chance to catch bad behavior before it acts.

As stated in the literature

Korbak and a multi-lab consortium, authors of a 2025 position paper framing chain-of-thought monitoring as a fragile-but-real safety opportunity; a touchstone in the debate over whether visible reasoning traces can be trusted for oversight.

Why it matters: It frames the central question of whether we can trust an AI's visible reasoning as a safety check, shaping how oversight systems are designed.

For example, this line of work treats an AI's written reasoning as a window worth watching for warning signs, while cautioning that a model could learn to close that window.

Heard on the show

“A multi-lab consortium published a position paper on it last year — Korbak and colleagues — basically calling it a fragile-but-real opportunity for safety.”
Episode 094 — Chain-of-Thought Monitoring Fails Across Languages, and Worst Where It's Needed Most

Mentioned in 1 episode

  1. 094
    Chain-of-Thought Monitoring Fails Across Languages, and Worst Where It's Needed Most

Related terms