I’d argue that tokenization is more analogous to having a losslessly encoded image where you don’t have access to the decoder. Humans don’t perform well on that one.