OpenAI researchers have proposed a method for predicting how often a new AI model will make mistakes once it is out in the world, according to The Decoder.
The idea is to estimate a model's likely failure rate before it ships, rather than only catching problems after users encounter them. The Decoder reports that the approach is meant to fill gaps left by standard safety testing.
That framing points to a known limitation of how AI systems are vetted today. Conventional safety checks can flag obvious problems, but they do not always tell developers how frequently a model will slip up across the messy range of real-world use. A forecasting method aims to put a number on that risk ahead of release.
The Decoder describes this as a proposal from OpenAI researchers, signaling work aimed at improving pre-launch evaluation. Beyond that, the available reporting does not detail how the prediction method works, how accurate it is, or whether it will be built into OpenAI's products.
Why it matters: if developers can reliably estimate how often a model will fail before it reaches the public, they can make safer release decisions and set clearer expectations — a meaningful step as AI systems are deployed into more high-stakes settings.