How AI agents shape experts’ judgment calls

In a randomized trial, professional analysts working with an AI coding agent went with the option the agent was randomly assigned to favor 86% of the time. We look at what this means for human oversight of AI-assisted work.

Read on anthropic.com