You're probably remembering that one METR study they did a little while ago. It was a bit limited, with an n=16 or something, and only focusing on specific, task-based coding solutions. But yeah, it showed that the coders were 19% slower at the task than those not using AI.