Analytic

In the error equation, overall error (mean squared error) equals bias squared plus noise squared, so removing a unit of noise is worth exactly as much as removing a unit of bias. Applied to agents: an ensemble of copies of one model drives the noise term toward zero and leaves the shared-bias term untouched, while a panel of independent people has the opposite profile. Neither profile is better in general. Which wins depends on which square is bigger, so the collective should measure both rather than prefer low variance, or diversity, for its own sake. This cuts against treating correlation as automatically fatal. The quotes are the authors'; the application is Rio's.

How groups think together