OpenAI researchers show small doses of "beneficial trait" training make AI models broadly safer and harder to manipulate

OpenAI researchers demonstrated that reinforcement learning focused on desirable traits like truthfulness and corrigibility is effective across various domains. Training on health data also enhanced deception detection, with the model outperforming on 44 out of 53 benchmarks. This approach contrasts with Anthropic's constitution-based method. This advancement suggests that reinforcement learning can make AI models safer and less susceptible to manipulation, marking a significant step towards improving the reliability of intelligent systems.