Human Evaluation

Deciding what to look at comes first

The most common failure is handing something off with just "check if this is good," with no rubric. One reviewer checks whether the sentences flow, another checks whether the facts are right, another checks the length. Pool the results, and you get numbers measuring different things all mixed together.

So the check gets broken into items. Did it answer what was asked, are the facts right, is it easy to read, is there anything offensive — each scored separately. Attach two or three worked examples to each item — "this counts as this score" — and reviewers line up far more closely with each other.

Building the rubric is itself half the evaluation. If you can't write down what a good answer looks like, that's often a sign you haven't decided what you're building yet.

Score it, or pick between two

Scoring a single answer is easy to set up but the scale wobbles. The same answer gets a 3 from one person and a 5 from another. The score even shifts depending on whether the answer looked at right before was good or bad.

So laying two answers side by side and asking which is better is used far more often instead. There's no absolute scale needed, which shrinks the gap between reviewers, and the same reviewer gives more consistent results across repeats. Pile up enough of these picks, and wins and losses convert into a ranking score.

The trade-off is that side-by-side picking takes more effort. Add more candidates and the number of pairs to compare grows fast. So instead of covering every pair, only the pairs that matter get picked out and shown.

What to do when people disagree

It helps to measure how often two people land on the same judgment looking at the same answer. A low number there doesn't mean the people did something wrong — it means the rubric is vague. Collect the cases where judgments split and look at them together, and the spots the rubric needs fixing come into view.

Having several people look at one item and going with the majority is common too. People's individual quirks tend to cancel out, so the result gets steadier. The identity of whoever wrote each answer needs to stay hidden, and the order the answers appear in needs to be shuffled — a habit of favoring whichever answer sits on the left turns out to matter more than you'd guess.

Slow and expensive is the real problem

The weak point of human evaluation isn't accuracy — it's speed and cost. You can't call in a room full of people every time a model gets tweaked a little. So automated metrics get used during day-to-day development, and human evaluation gets reserved for the moments that actually matter.

These days, having a language model do the grading is widely used too. It's fast and cheap, but it has its own quirks that differ from how people judge, so whether it can be trusted comes down to how closely it lines up with human evaluation results. That's exactly why, no matter how much automated grading grows, human evaluation never quite goes away.