Human Evaluation

What is human evaluation?

Some things are hard for machines to measure — whether an answer "reads well" or "offends someone". Human evaluation puts real people in front of AI output, asking them to score it on several dimensions and write comments, to judge whether it's actually good.

Why can't we skip the humans?

Quality is subjective
"Helpful", "creative", "appropriate tone" — these are hard to measure precisely in code. Human judgment is the most direct.
Safety needs a human eye
Discrimination, dangerous advice — these red lines still need people to catch them today.

How is it usually done?

Scoring
Evaluators rate outputs on helpfulness, accuracy, safety and so on, with a 1–5 scale or good/neutral/bad.
Side-by-side ranking
Put two models' answers next to each other and ask which is better — the same preference ranking that powers RLHF.
Written feedback
Not just scores, but "why it's bad and where" — that written feedback is gold for improving the model.

The costs, and the workaround

Human evaluation is expensive, slow, and shaped by evaluator bias. So teams usually pair it with automation: automated evaluation runs at scale for quick coverage, while humans watch quality and make the final call.

Bottom line: human evaluation is the AI's "human examiner" — the things machines can't measure, people decide.