a builder's codex
codex · operators · Daniel Kahneman · ins_noise-vs-bias-judgment-quality

Noise is at least as damaging as bias — and most orgs have no instrument to even see it

By Daniel Kahneman · Nobel laureate; Princeton emeritus; co-founder of behavioral economics · 2021-05-18 · book · Noise: A Flaw in Human Judgment

Tier A · TL;DR
Noise is at least as damaging as bias — and most orgs have no instrument to even see it

Claim

Bias is the average error across many judgments — a systematic skew. Noise is the unwanted variability between judgments that should be identical: two interviewers rating the same candidate differently, two PMMs scoring the same launch, two doctors diagnosing the same patient. Noise is at least as damaging as bias, far more underappreciated, and invisible without measurement.

Mechanism

Decision quality has two failure modes. Bias is consistent error in one direction — easier to spot because patterns appear across many decisions. Noise is unstructured error — it cancels in the average, which is why org-level metrics make it look like nothing is wrong. Two PMs presented with the same competitive intel will write different positioning recs not because either is biased, but because mood, recent context, anchoring, and order-of-evidence-encountered shift each rater independently. The fix is structural: independent written ratings before group discussion, rubric-based scoring, mechanical aggregation, and noise-audits that measure inter-rater variance directly.

Conditions

Holds when:

Fails when:

Evidence

"Bias is the average error across many judgments (always too high or too low), but noise is the unwanted variability in judgments that should be identical."

"noise is at least as damaging as bias and far more underappreciated"

— see raw/expert-content/experts/daniel-kahneman.md lines 16, 18.

Signals

Counter-evidence

Standardisation can flatten edge cases that experts would correctly handle as exceptions — the rubric becomes a ceiling on quality, not a floor. Gary Klein's Sources of Power argues that in pattern-rich expert domains (firefighters, ER doctors), the noise is correct because the experts are integrating cues a rubric cannot capture. Noise reduction is therefore a means to a quality end, not a quality goal in itself.

Cross-references

Open the interactive view → View original source → Markdown source →