Human Evaluation
Human evaluation is a person scoring answers against a written rubric — the gold standard for "was this actually good?" Use it to define quality and to check cheaper graders, not to label every production reply.
Explain like I'm new to AI
A thumb up/down hides why. A rubric splits the job: grounded, helpful, safe, on-brand. Raters score each axis (or pick A vs B, which humans find easier than a lonely 1–5).
A support answer. Score each axis — thumbs hide which part failed.
1–2: Invented a policy. Sounds confident.
4–5: Answers the actual ask with the next action.
4–5: Refused or deferred; no PII, no overclaim.
Rubric average 3.7 vs a thumb 👍 'looks fine'. Same answer; only the rubric tells you to fix grounding, not tone.
You cannot hire humans for every chat. You hire them to (1) build the gold set, (2) write the rubric, (3) calibrate LLM judges, (4) sample live traffic. Domain work (legal, medical, finance) needs experts, not random crowd workers.
Mental model
Olympic judging: published criteria, more than one judge, disagreement is a signal the rule was vague — not a reason to throw out scoring.
How it works
- Write the rubric before labels. Define 1 vs 5 with examples (the visual above).
- Train raters on 20 items together until they mostly agree.
- Measure agreement (if two people constantly clash, the task is underspecified).
- Prefer pairwise for close calls ("which summary would you send?").
- Sample: 100 well-chosen items beat 5,000 tired clicks. Watch fatigue — quality falls after long sessions.
- Use those labels as the source of truth for judges and for CI.
Real-world example
Two raters score "helpfulness" all over the map. You add one line: "Helpful means the user can act without another lookup." Agreement jumps. The model wasn't random; the spec was.
Technical explanation
Inter-annotator agreement (Cohen's κ, Krippendorff's α) is how you know the rubric is real. Low agreement → fix instructions, don't average noise.
Bias: raters punish unusual phrasing, prefer their dialect, get primed by the last bad answer. Shuffle order. Hide which model wrote it.
Cost: humans are slow and expensive. That is why the next lesson exists. Humans remain the calibration, not the production mill.
For agents: grade the trajectory (wrong tool, skipped confirm) not only the final sentence.
Common mistakes
"Looks good" from the PM on five chats. That is a demo, not human eval.
- Rubric written after you already picked a winner (you retrofitted taste).
- Crowd labels on medical advice.
When to use it
- New product, new rubric, new judge, new high-stakes slice.
- Periodic audit of live traces (even 50/week).
When NOT to use it
- Grading every token in production. That's what unit checks + judges are for — after humans set the bar.
Alternatives
- LLM-as-a-judge (next) for volume. Exact match / schema tests when the job isn't open-ended. Online task-success when you can instrument the outcome.
Quick quiz
Related concepts
- LLM-as-a-Judge — LLM-as-a-judge grades outputs against a rubric at scale. Calibrate it on humans, and watch verbosity, position, and self-preference bias.
- Hallucination and Safety Evaluation — Hallucination eval checks groundedness vs context, not "sounds true"; safety eval checks refusals, leaks, and overconfidence. Score them separately.
Last reviewed: 2026-09-04 · Written by ByHeart AI · Reviewed by ByHeart AI