New AI Training Method Makes Models Broader, Safer, and Harder to Manipulate

Researchers at OpenAI have found that training AI models on realistic scenarios with desired behavioral traits can make them safer and more helpful across various domains. By testing specific traits like truthfulness and epistemic humility, the team discovered that good behavior can transfer to unfamiliar domains and even improve performance on unrelated benchmarks. This approach is fundamentally different from Anthropic’s constitutional method, which relies on explicit values documents. The OpenAI method shows promise in making AI models more resistant to manipulation, but further research is needed to compare the two approaches.