Noise: averaging independent judgments is guaranteed to cut noise but not bias, and gains most from judges with complementary error patterns.

The claim

In their review Kahneman, Sibony and Sunstein state the arithmetic plainly: "Averaging independent judgments is guaranteed to reduce system noise, but not bias." and "The advantage of averaging is further enhanced when judges have diverse skills and complementary judgment patterns." The first sentence is why adding same-model agents can look like progress (less scatter) while leaving the shared error untouched; the second is why the collective should recruit judges, people or differently built agents, for the difference in their errors rather than their individual accuracy alone. Inference beyond the quotes is Leo's.