While particular battles are "set up" in HL, the behavior of the troops in particular environments (such as hiding behind cover and simultaneously chucking a grenade at the rock behind which you're hiding) clearly falls into the same category as what FEAR AI does.
All battles in Half-Life are set up. They are all scripted.
How is that different from what individual AIs do, say, in Men of War? They pick up weapons while going about other tasks, too (though they could be smarter about it).
Implementation. Instead having to tell every possible actor that Weapon X can be picked up, you make Weapon X possible to pick up.
Well, strictly speaking, if you were in the exact same situation, it would still do the same thing, unless it's randomized (little incentive to do that). Of course, the players never quite does the same thing, so the reaction may be different.
Sure, if you play pixel-perfect every single time.
The ability of the AI to do these things would have been revolutionary...about 15 years ago, if not 20. The question is how well it does these things, not whether it should follow a hard script. However, all they describe is an AI that solves a bunch of optimization problems of some sort or another, or runs a "pathfinding" algorithm of sorts through the action space. The end result is still fundamentally a "script," just a more sophisticated one. There's nothing revolutionary here.
Didn't say it is, per se, revolutionary (except it is, since it allows behavior without scripts, and is a major break-through in AI research). However, it allows for more believable behavior with less effort by the developers. A scripted AI needs to be told what the map looks like, what objects can be used and how, and you have in general a harder time to get the behavior you want. Not to mention that a goal-oriented approach is much more flexible. Instead of the AI having to run throug ha state-machine, as well as the developers setting entry and exit points (for example: "build 20 more INF divisions, put attack on Finland on hold"), the AI evalutes what it has, looks for what tools are needed to reach a given goal, and sets about building what it needs.
For example, in HoI2 the AI cancels Operation Seeloewe, if it loses a given number of transports (I think 2/3, but it doesn't really matter). Doesn't matter if it has the IC or the time to rebuild it, Operation Seeloewe is canceled. The state-machine is finished, and the AI is being told to go do something else.
Similarly, Operation Barbarossa gets canceled, if the AI doesn't reach a given threshold.
And look what happens if you break the expected flow of the game. The AI is completely lost. It isn't able to exploit weaknesses of the player (the STRAT bombers just keep coming, despite them being shot down like flies. Patrol a port area with plans on Naval Strike, and the AI still uses the port, instead of switching to another one).
Of course, with enough scripting you can solve all of that. If you give the AI goals, and actions that can be used to satisfy these goals, you don't have to anticipate the SU declaring war on AI-China, since the AI will still be able to cope. In an alliance, the AI could actually research useful stuff, depening on what the players / other alliance members do, instead of stupidly following the list of preferred research. No nuclear research, if the Germans are knocking at the Kremlin's door.
No need to block subs for the AI, since it can't make proper use of them.
Once the AI is somewhat aware of the game-world, it could set about convoy raiding in an efficient manner, instead of using uber-stacks.
No reinforcing of beachheads it can't exploit, instead it'd change the thrust of an offensive, or retreat from a stale-mate. Oh, it can't since it isn't able to find out how many transports it needs, since the script hadn't foreseen this happening.
In short: You generalize your scripts as much as possible, and make them re-usable. GOAP is to state-machines what OOP languages are to procedural languages. Both do the job, one's just easier to handle and produces better results.
EDIT: While we're at it, don't you find it highly ironic that both Demigod and ETW, the only strategy games on that list, have been lambasted for poor AI? I'm not saying that the paradigm is to blame, just pointing out that throwing buzzwords around doesn't save anyone from coding the damn thing intelligently. It particularly odd since Demigod seems to be a fairly straightforward RTS (in fact, much more primitive in its environments and features than a lot of other stuff out there, such as Company of Heroes, or Men of War).
Demigod's geared towards multi-player. The AI there is a tool to learn the game itself.
Besides, in highly structured environments a state-machine can behave better than anything else (after all it's been used for ages now). Once dynamism comes into play, things get interesting. In the Chinese meaning of the word.
Ever played Sims? That's a GOAP-ish AI at work, too. And Sims work rather well if let to their own devices (that they aren't able to do it perfectly is the point of the game).
EDIT2: In fact, I recommend typing in something like "empire total war terrible ai" in Google, some hilarious results here (amusingly enough some of them relate stories of terrible AI in FEAR 2, as well). I can comment that Fallout 3 AI was acceptable, but it wasn't anything to write home about.
Half-Life 2: Grab a barrel. Keep it in front of you, and make sure it covers your enemies. Enjoy your poor man's god mode. The AI won't act, at all, if the player can't see the actors. Valve claims it is out of fairness to the player. The result is that a) you can't use an interesting tactic and tool properly (the grav gun to provide cover for you while you advance), and it is hilarious that you can basically walk up to a guard that's staring in your face (well, the barrel infront of it, anyway), and it won't do a thing.
Dark Reign: The AI was called revolutionary because it'd learn the player's strategy and adapt to it. Didn't work out that well, you could still tank-rush it, even after hours of play.
C&C1: The AI wasn't able to cope with a player blocking the entrance of the AI base with a concrete wall (sandbags didn't work that well, since tanks could run it over). The devs didn't anticipate that, and didn't include this eventually in the scripts or the triggers.
Of course it depends on implemntation, and design how well something works. There is no silver bullet. But games as complex as Paradox's games are push the limits of scripted events to the breaking point. Quite often beyong it, too.