Autocorrect(and predictive text) has reached an interesting milestone, good enough that you mainly rely on it, but occasionally it distorts the data you intended to send. Disengaging the technology isn't easy, or even understanding field types it shouldn't correct(everyday my work VPN auth window tries to autocorrect my username). There is so much context that is unavailable to the computer, and there is no good way to provide it.

I think it has interesting implications for self-driving, since as far as I can tell we're at an earlier but similar stage in its development; works for simple situations, likely to drive you into a wall for no reason. We build so much trust in these systems, we stop paying attention to what they are doing, and then pay the price when it behaves in an unexpected way to the user.