This is one of many examples why claims regarding near- or super- human performance in AI papers need to be taken with a good amount of salt. CIFAR-10 is great for experimenting with your algorithms, but it's a horrible dataset to do any kinds of human-to-machine performance comparisons.