OpenAI simulates real-world tests to catch AI misbehavior

OpenAI unveiled deployment simulation, a method that feeds unreleased AI models real-world chat data from already deployed AI. This prevents models from detecting they are under evaluation. The technique aims to reveal undesirable behaviors like lying or harassment before public release, improving safety alignment.