How does LLM-as-a-Judge evaluate another AI's response?

An AI teacher grading an AI student — with a marking scheme.

Picture it first

Your teacher can't check 1,000 essays alone, so she trains a senior student as an assistant marker and gives them a clear marking sheet: '2 marks for the right answer, 1 mark for showing working'. She checks a few of the assistant's marks herself to make sure they're fair — especially because the assistant likes long essays a bit too much.

Now the same idea, in simple QA words

Checking thousands of AI answers by hand is impossible, so we use another AI as the examiner. We give it the question, the answer, the correct information and a clear marking scheme (e.g. 5 = fully correct, 1 = wrong). It returns a score and a reason. But the examiner AI has habits too — it may prefer longer answers or the first answer it sees — so we first compare its marks with human marks before trusting it.

How the story connects to the real thing

In the storyIn the AI system
Teacher with 1,000 essaysTeam with thousands of AI answers
Senior student as assistant markerLLM-as-a-Judge
The marking sheetThe scoring rubric
Teacher re-checking a few marksComparing judge scores with human scores
Assistant liking long essaysJudge bias toward long answers

Real-world example

A team scores 2,000 chatbot answers nightly. The judge prompt asks for JSON: {correctness: 1-5, groundedness: 1-5, reason: '...'}. Before trusting it, they had 3 humans label 100 answers; the judge agreed with humans ~88% on pass/fail. They also swap A/B order in pairwise tests to cancel position bias.

Step by step, in plain words

  1. Write a clear scoring rubric with examples.
  2. Give the judge AI: rubric + question + answer (+ correct info).
  3. The judge returns a score and a short reason.
  4. Check the judge against human scores on a sample.
  5. Fix the rubric until judge and humans mostly agree.
  6. Use the judge on all answers; re-check now and then.

How a tester tests it

  • Calibrate: measure judge-vs-human agreement before trusting scores.
  • Bias tests: swap answer order; pad an answer with fluff — does the score change?
  • Pin the judge model version; a judge upgrade changes all your historical scores.
  • Check judge consistency: same input scored twice should match.

Try it yourself

1) Write a 1–5 rubric for 'helpfulness of a customer-support reply' with an example for each score. 2) Give the rubric and 3 replies to a chatbot and compare its scores to yours.

Interview angle

“How do you know you can trust an LLM judge?”

I validate it like any test instrument: compare its scores to human labels on a representative sample, check agreement, test for position and verbosity bias, keep the judge model pinned, and re-calibrate when the rubric or model changes. An uncalibrated judge is just another unverified model.