OpenAI researchers want to predict how often AI models will fail before launch

In three linesOpenAI researchers propose a method to predict how often a new AI model will make mistakes after release. This approach could fill gaps left by standard safety testing.

Summary generated by Claude — human-verified