WildChat-based simulation predicts OpenAI production misalignment rates within roughly 3x

OpenAI's alignment team found that re-generating 100,000 WildChat conversations with five recent OpenAI models predicted production failure rates across four orders of magnitude, with 95% of predictions within 1.04 orders of magnitude.

The approach was weaker for agentic misalignment categories, where errors were about 37 times larger, and it still held roughly without access to chain-of-thought reasoning, with mean multiplicative error rising from 3.6x to 4.0x.