Driving, diagnosing, and translating are not good comparisons, because none of those are adversarial.

The problem here is that adversaries are adjusting to whatever measures are put in place. In that respect it might be more like winning at Chess or Go. Which computers can do, but it’s decidedly non-trivial.

And here’s the kicker: If we postulate some hand-wavium big-data/machine-learning/AI that can detect bots and adjust when the bots evolve, why can’t we also postulate some hand-wavium big-data/machine-learning/AI that can run bots and evade the bot detection?

Driving often is adversarial, and no AI can actually drive yet (outside of some very limited circumstances on a small number of roads).

It's probably not that they can't classify fake accounts but that doing so at a scale requires too large resources and it's more economical to employ only simpler techniques that weed out most of the bots and live with the remaining few.

Then if they also happen to anyhow benefit from the existence of those fake accounts then it becomes even harder to justify costly filters and leads to a sufficient-enough-to-not-look-obviously-weak grade filtering