Noise and bias are separate, independently measurable errors

The book's foundational distinction is between bias, a consistent, directional error, like a scale that always reads two pounds heavy, and noise, an inconsistent, undirected scatter of errors, like a scale that reads differently each time you step on it even though your weight hasn't changed. Total error in a system of judgments can be mathematically decomposed into both components, and reducing one doesn't automatically reduce the other.

The authors illustrate this with a target-shooting metaphor: biased shots cluster tightly but off-center; noisy shots scatter widely around the correct center; a system can suffer from either, both, or neither. Noise can exist even when the average of many judgments happens to land near the correct answer, because errors on either side of the target can cancel out in aggregate while still representing serious inconsistency for any individual case.

This matters because organizations that audit only for bias, checking whether outcomes skew systematically in some direction, can completely miss substantial noise sitting underneath an unbiased-looking average.

Takeaway: an average judgment can look perfectly accurate while individual judgments underneath it vary wildly and unfairly — always check for noise separately from bias.