Comment by skybrian

They are being trained to try lots of unlikely alternatives and to be persistent. This often works well when searching for security bugs or counterexamples to famous math conjectures.

But maybe it doesn't work so well when caution is required?