OpenAI API capability update: teaching models to express uncertainty in words

AI Impact Summary

This CAPABILITY change enables language models to output explicit uncertainty in their responses, enabling end users to gauge reliability without leaving the conversation. Implementing this safely requires prompts or system messages that phrase confidence levels consistently and guardrails to avoid overstatements in high-stakes domains. Technical teams should plan UI elements to display uncertainty, metrics to tune the frequency and tone of uncertainty signals, and testing to prevent warping user trust. The evolution will improve decision quality in chatops and advisory assistants, but will necessitate monitoring and policy controls around how and when uncertainty is disclosed.