Things that are unsurprising probably are frequently mentioned in the training data. Just based on that you would expect that language models have an easy time judging how novel information is, you just have to get it to use that information.

I would expect GPT3 to be able to do that, but I don't think this is GPT