Three times in four, the AI judge picked the first answer.

transcript

You ask an AI to judge two answers. Which one's better? Three times in four, it picked the first one. Not the better one.

Researchers showed judges two similar answers. Then they swapped the order. Same answers, same question. A fair judge keeps its verdict.

GPT-4 kept its verdict in 65 percent of cases. The better of the two. Claude-v1? Just 23.8 percent. Most of the time, its verdict changed when the order did. It favoured the first answer 75 percent of the time.

Here's the catch. Those were 2023 models. Today's judges have their own biases. Position, length, even liking their own answers. You don't know until you test.

So ask the judge twice, with the order swapped. A win counts only if it holds both ways. Otherwise, it's a tie. Then check its verdicts against human labels, before it gates a release.

A judge is a measuring instrument. Check it first. A win must survive the swap.

Full breakdown in the lecture.