Differential productivity gains and defect escape rates in AI‑assisted AI R&D

Quantify how productivity improvements from AI tools differ between AI safety teams and other teams within AI research and development organizations, and measure how often AI‑introduced errors in AI R&D outputs escape or could escape human review.

Background

Beyond measuring the extent of automation, the paper highlights a lack of data on AIRDA’s consequences, including whether safety research keeps pace with capabilities and whether oversight processes remain effective. Understanding differential productivity is critical for assessing whether defensive and safety work is being accelerated sufficiently relative to offensive capabilities.

Measuring the rate at which AI‑introduced defects evade review is essential for tracking oversight effectiveness and the oversight gap, particularly as automation may increase the volume and complexity of outputs requiring scrutiny.