How AI agents shape experts’ judgment calls

Analysts evaluated a farm program with the help of a coding assistant. The researchers randomly assigned the assistant preferences about four choices, including which farms counted and how to weight them. Each choice could change the estimated benefit.

Among participants who used the assistant, the analysts’ methodological decisions agreed with its assigned preferences 86% of the time. Its direct influence on the final policy recommendation was smaller. The experiment locates influence in the preliminary decisions that determine what number a board sees.