Teaching models to express their uncertainty in words

OpenAI has published research on a method to make language models communicate their uncertainty in natural language, according to the OpenAI Blog. The approach involves training models to produce verbalized confidence levels alongside their answers, rather than relying solely on numeric probability scores. This could help users better interpret the reliability of model outputs in conversational AI. The technique uses a calibration dataset where models learn to match their stated confidence to actual accuracy. For builders integrating AI into workflows, this addresses a common pain point: knowing when to trust a model's response. By teaching models to say 'I'm not sure' or 'I'm confident,' the output becomes more transparent and actionable, especially in applications requiring high reliability, such as code generation or data analysis. The research is still early-stage, but it points toward more trustworthy AI interactions.