I've actually started trying to see how wrong the newer Capchas from will let me be on purpose. Either by not selecting all of them or picking wrong ones. They let you through a lot of the time.
My working theory was that companies like Google were using the capchas mostly to generate AI data, so only a few of the images on any given test were actually already labeled. Any of the other images (particularly the really grainy ones) would accept any answer because they were genuine classification questions.
Reading this article, I wonder if it's not even that -- that companies like Google are assuming, "you're not going to get everything right, so we'll give you some leeway."