(Subscription required) Early this summer, I ran an experiment in which I gave ChatGPT-4 and ChatGPT-5 only the oral argument transcripts from the court’s last 25 cases and asked them to predict the outcome and the vote. Here were the results.
