Emphasis mine.
They use an AI trained on PvP player behaviour, so the NPCs jump around, respond to teammates being injured, try to catch you off-rotation, etc.
In my opinion, it wouldn't work for bosses though, as part of the fun is exactly about understanding the tells and patterns. If you want to fight against intelligent enemies, just play PvP! I think if you play without addons like BigWigs/DeadlyBossMods you'll have much more fun, if what you're looking for is unpredictability.
Plus, would there really be that much of a difference between a machine-learned approach and a decision tree when every action the boss can perform needs to be pre-programmed anyway, and have predefined ranges/optimal uses? Maybe a different game could combine that with something like Magicka, where there are spell "recipes" which can be combined, so the AI would have more options to play with.
Another example is things like Dark Souls, the bosses are so engaging to its players _because_ you can learn to understand how they respond to your positioning. If they were a moving target, there would be no point in doing anything but learning the animations for each attack so you can dodge, so in the end they would be _less_ intellectually engaging.
The AI could be unaware of its tells or that it's telegraphing moves. Even more wild would be if the AI could be aware, so that maybe it tries to cover the status light or hide the weapon from view while it charges up. That in itself would be a tell.
That seems like a reasonable first step imho.
So we're stuck in a kind of AI limbo.
Efforts exist in beating humans in specific games like chess; these games are often symmetric, i.e. all players compete on fair ground, and the AIs aim to replace human players. But most games don't want this! Most games seek theatrical asymmetry in computer-controlled agents. Most games want the illusion of cleverness without actually being clever, because most players don't want their gaming experience interrupted by undefeatable killer robots. Most games are power fantasies.
I posit that what we actually want is simply some programming pattern for managing complexity in autonomous theatrical agents. Some way to organize the creative will of the game designer easily enough that it democratizes independent developers to craft interesting behaviors.
I've got my own ideas but they're as of yet unproven.