OpenAI explores public chat data to predict real-world AI misalignments

OpenAI has published a blog post exploring whether public chat data can predict real-world AI misalignments, specifically targeting the gap between controlled benchmark tests and actual deployment behavior. Traditional evaluations rely on hand-written or adversarial prompts, which can be narrow or easily recognizable by models. OpenAI's alternative approach, called 'Deployment Simulation,' originally used internal production data to sample real user interactions and check for undesirable model behavior, including rare pathologies. However, because production data is private, external groups like AI safety organizations, academics, and independent researchers cannot access it.

To address this, OpenAI is now testing whether a publicly available dataset—WildChat, which contains real user conversations with chatbots—can serve as a substitute for internal production data. The goal is to allow external evaluators to run similar simulations and detect misalignments without relying on labs' proprietary data. If successful, this method could close the gap between benchmark results and real-world consequences, making frontier AI safety evaluations more transparent and accessible. The research is shared on the OpenAI Alignment blog, emphasizing the need for diverse evaluation sources to ensure models behave safely in high-stakes settings.