Fun concept. Unsurprisingly a lot of these are common jokes 'plagiarised' directly from the training data. I wonder if there's a way to integrate a search tool to rule out those with near-exact matches on the web.

It would be hard to do well, but something basic could probably at least make a dent in forcing creativity out of the models.

Edit: better yet - manually curate a few dozen unseen 'first halves' of jokes and have the models complete the joke