>But for which of these tasks can ML models outperform us at given the same amount of data as we have?

The brain comes pretrained for a lot of things tho. Babies react to human faces. Pareidolia is simply our pretrained brains "overfitting" random noise, interpretting it as faces. Because, most probably, humans are very good at recognizing human faces. It's a great question really, because it makes us think about the nature of our own intelligence.

A better comparison could be made, maybe, by selecting tasks at which a human would be not "pre-trained".