순서

The biggest obstacle for AI in real-world settings is that most human tasks aren't nearly as tidy in their rules and boundaries.

(A) Those "gray areas" (like driving on busy city streets) are where AI tools often falter. Despite years of expert forecasts of their imminent arrival, we don't yet have self-driving cars that can safely replace human drivers, other than in controlled settings.

(B) AI performance in these areas can still improve, as the size of AI models grows, and they are trained on more human driving data. But they will improve only by mirroring more of the patterns found in our intelligent performances, as we cope with a world that no data set can ever contain.

(C) That's why some driverless taxis created chaos on the streets of San Francisco. Despite years of optimistic predictions, many AI tools cannot yet be trusted in open-world environments, where the human stakes are high, and failure means real, permanent loss.