One of the interesting things we're finding as we start to coax machines to do human jobs is that humans tend to not accept a machine unless it's much better at the job than a human. For example, on many computer vision or NLP tasks, the current state of the art is about as good as most humans.

The problem is that when they're wrong, they're wrong in ways that humans wouldn't be, and the wrongness appears to be "broken" to us and we overweigh these incorrect responses because they're both wrong and not human in their wrongness.