liveBook · Manning

Noise

As explained in AI Model Evaluation

Noise is random or irrelevant variation in an observed evaluation result that does not represent the product effect of interest.

Noise is random or irrelevant variation in an evaluation result that does not represent the product effect of interest. It can come from user behavior, sampling, unstable inputs, day-of-week effects, novelty, or an inconsistent evaluator. section 6.5.4 section 8.1.1

In an A/B test, noise makes observed effect sizes swing and increases false positives and false negatives. Small effects, rare conversions, long feedback loops, and unstable metrics make the problem worse; larger samples, stable metrics, and appropriate evaluation windows help separate signal from randomness. section 6.5.4

Noise commonly appears in three evaluation problems:

  • Weak metrics elevate irrelevant movement into apparent success. A higher click-through rate can reflect attention-grabbing titles rather than improved satisfaction, consumption, or retention. section 8.1.1
  • An unstable judge can make model comparisons meaningless. If judge variance exceeds the score gap between variants, the apparent winner may only have received a favorable run. section 10.4.1
  • A noisy gate can approve or reject a model according to the sampled run. A pass rate that moves across the decision threshold does not reliably measure quality. section 10.4.1

Movie recommendations

A recommender increases clicks but reduces long-term viewing satisfaction. The click increase may be attention noise or a weak proxy for product value rather than evidence that recommendations improved. Compare metrics aligned with the product outcome, and inspect whether the observed lift persists across suitable evaluation windows. section 8.1.1

When an evaluation result changes unexpectedly, triage the measurement system before blaming the model. Compare canary scores with a stored baseline, check data and prompt or rubric changes, repeat samples, and swap evaluation order. If instability or order sensitivity remains, treat the result as judge noise; if those checks are clean and only the model changed, the regression is more credible. section 10.6.2 section 10.4.1

How AI Model Evaluation covers it

How deep it goes
Explains sources, consequences, mitigations, and triage checks, but does not develop formal statistical methods.
Examples
Uses movie recommendation CTR, variant score gaps, unstable pass-rate gates, and canary evaluation triage.
What you can do after
Diagnose noisy evaluations with baselines, repeated sampling, order swaps, stable metrics, and appropriate evaluation windows.
How it connects
Connects noise to A/B testing, signal, metrics, retention, model evaluation, and LLM-as-a-judge stability.

Also explained in

Compare how 16 books explain noise →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Model Evaluation. The section links point to the parts of the book it draws on.