True, UT had reasonable AI but it did always seem to perform much better on deathmatch where it could just run around shooting things than on CTF maps where you could quite easily camp outside of your base and just snipe it when it tried to come in.

Of course AI is getting better now that we can offload stuff onto other cores without tying up the game loop so much.

I don't know to what degree the UT bots qualify as AI players... their strategies are coded into the level, not the bots. (disclaimer: I read guides to making UT levels, but never actually made one. Details may be misremembered. However, the fact that two levels with identical topography might be treated entirely differently by the same bots (owing to different bot-path placement in each) doesn't leave me filled with awe at the bots' decision-making.)

Placing paths manually or having them auto-generated by the map editor is standard practice. It's mostly because it would be expensive/silly to calculate at runtime -- calculate it when it changes, save it in a format bots can parse; the bot AI still makes decisions based on that path. How is that different from first calculating it?

Can the paths be usefully autogenerated (at whatever computational cost)? Or do they have to be done manually? The latter puts us a good deal further away from "real" AI.

(Again, to the best of my knowledge) the paths UT bots follow through a map are laid down by the author of the map. You can go into existing maps and change the bot paths.

I don't know what happens if you put bots in a level that doesn't have bot paths, but it wasn't advised.