I think this sort of thing will improve once AIs are trained using the human's enjoyment of the game as the primary utility function. This could be as simple as asking users to rate the preceding game at its conclusion.

It could be as subtle as using the player's apm or other activity metrics as a proxy for enjoyment, if correlations were shown in play testing labs. This could allow for more training data, and less annoyance to users, who might be reluctant to provide manual ratings.

The next step after that is adaptive AI, where the computer learns what you find enjoyable personally.

I give it ten years before this is how every game works.

Edit: Google has said that negative signals are extremely important to its search. A user bypassing the #1 result for the #2 is hugely bad for #1, but only a mild boost for #2. Similarly, perhaps minimizing "ragequit"s is a valid objective!