"Humans are not great at exceptions, but we are still, at least for now, better than machines"

An interesting statement made by Erik Brynjolfsson during his AI, Science and Society Conference hashtag#AIActionSummit at École Polytechnique/Institut Polytechnique de Paris.

The "at least for now" part.

Erik made that point by showing a Pareto-looking curve of question frequency from a customer support chat, with the most frequent questions generating a lot of data that can be used by machine learning to provide an answer while the least frequent questions are better addressed by a human. The fact that generalization breaks down for way "out-of-distribution" questions is an obvious weakness of statistical machine learning and an opportunity for humans. But that's a statistical learning perspective. As machine reasoning improves, the long tail of questions may prove to be assailable.

link (around 6:33:00)