Eliezer Yudkowsky argues alignment training can hide dangerous behavior

DoomBench assesses “Eliezer Yudkowsky argues alignment training can hide dangerous behavior” as evidence moving toward doom, with magnitude 32 and confidence 58 out of 100 in the safety and alignment category.

The DoomBench assessment of “Eliezer Yudkowsky argues alignment training can hide dangerous behavior” is based on reporting from The New York Times and records the editorial rationale, source quality, attribution, and revision history.

DoomBench summarizes “Eliezer Yudkowsky argues alignment training can hide dangerous behavior” as follows: In a full interview with Ezra Klein, Yudkowsky argued that optimizing models against visible bad behavior can select for behavior...