Where do you think the training data for AI judges would come from?

evaluations of historical judgements made and what the consequences were.

How to define good/bad consequences becomes the problem then, but thats the nature of it.

Such evaluations are useless as a training set because they inherit the biases of whoever wrote the evaluations. Of course, if fairness isn’t a desired property, feel free to carry on.

It's AI judges, all the way down.

I, for one, welcome our new AI overlords.

And who gets to decide what training data/processes the judges use?

AI Training Specialists who are appointed by the party in power, of course! So different!