OpenAI researchers say that training an AI model on a few desirable behavioral traits can make it safer across the board — and harder to manipulate.
According to The Decoder, the researchers used reinforcement learning to reward traits like truthfulness and corrigibility (a model's willingness to be corrected or shut down rather than resisting oversight). The notable finding is that these gains transferred across domains, rather than staying narrowly tied to the examples the model was trained on.
In one example reported by The Decoder, training the model on health-related data also improved its ability to detect deception — a skill the training didn't explicitly target. The Decoder reports that the resulting model scored better on 44 of 53 benchmarks, suggesting the broad safety improvements didn't come at a steep cost to overall performance.
The Decoder frames OpenAI's method as distinct from the "constitutional" approach associated with rival lab Anthropic, which guides model behavior using a written set of principles. OpenAI's technique instead reinforces the traits directly through learning signals.
The practical promise here is efficiency. If a small amount of targeted trait training produces wide-ranging safety benefits — and makes a model more resistant to being tricked into harmful behavior — labs may not need to anticipate and patch every failure mode individually.
The research has been presented by OpenAI's own team, so independent verification of the cross-domain claims will matter before the approach is treated as settled.
Why it matters: as AI systems take on more consequential tasks, methods that make them reliably honest and correctable — without crippling their usefulness — are central to keeping powerful models trustworthy.