How we detect frontier model misalignment
OpenAI describes how it monitors frontier models, investigates concerning behaviour, and tests the limits of its safeguards to detect misalignment.
Research
OpenAI Alignment
OpenAI's alignment research blog. alignment.openai.com
Focus All
Topic Control + Evals
Source OpenAI Alignment · 2
Date Any time
1 item
1 item
Beyond g-mean²: complementary metrics for monitorability
Proposes complementary metrics, pair-matched comparisons and post-control risk, to evaluate how monitorable models are, supplementing the g-mean² score.