True, but I think Qworg was instead suggesting that humans are good where the confidence/accuracy of robots drops off. Machine learning tasks usually get to 70% good enough pretty quickly, but after that each percent gain costs more and more. There's a point where humans can complement robots when the robots are not yet good enough to solve the problem.
For simple tasks. Identifying thousands of different products by sight alone in infinite numbers of configurations at varying light levels/obstruction? Humans are best.
