OpenAI tests if good behavior can generalize across domains

openai researchers tested whether reinforcing good behavior via reinforcement learning in specific domains generalizes to others. they trained on traits like truthfulness and fairness across healthcare, law, and education, then tested on 53 benchmarks. the model improved on 44 evaluations covering deception and reward hacking, showing that positive behavioral patterns transfer broadly. additionally, the 'beneficial trait' model showed 'selective persistence', resisting adversarial prompts and harmful fine-tuning while remaining steerable for helpful tasks. this contrasts with anthropic's constitutional ai, as openai relies on empirically measurable traits rather than a written constitution.