Assessing AI output — and judge bias, noise and subjectivity
How we assess whether an AI answer is correct — programmatic checks, judges, and numeric scores. Ground truth, AI judges, and judges of judges.
▶
What we cover
- From labeled classifiers to free-text GenAI: what you lose and gain
- Multi-criteria LLM judging, and how an "overall" score is a chosen combination
- Why mapping 1–5 scores onto 0–100 can flip which system looks better
- Judge bias and noise compound with sampling noise
- Playground intuition: redraw samples and dial judge noise to see measured quality drift
Understand · report · reduce — bias and noise around a score
- Understand — the judge is an instrument with two defects: noise (random wobble, averages out) and bias (systematic offset, more data can’t fix it).
- Report — give agreement with trusted labels (e.g. Cohen’s κ) and show the rubric, never just the headline number.
- Reduce — noise: average repeats, lower temperature, sharpen the rubric. Bias: calibrate against a trusted slice.
Skills you unlock
- Decide when exact scoring is enough vs. a rubric judge
- Treat the judge as part of the measurement system
- Make score-aggregation and mapping choices explicit
- Separate sampling noise from judge bias and noise
- Report judge setup and uncertainty with the score
Hands-on
No install, no code — two judge types, two defects: turn the dials and feel the difference between bias and noise.
Scope note: this lecture treats the judge as an instrument with measurable defects. Subjectivity — “the judge has opinions” — is the second half of this lesson.