The first problem is, AI is expensive. In terms of computing power, not money, although both usually apply.
The PS4 is seriously challenged as it is. Not its GPU actually, which is still better than in most mid-range laptops (like mine, I’m playing the game underspeced), but its CPU has lost the race long ago. The dual-core I7 that is by now practically standard even in the low-price segment of laptops runs circles around it, and the CPU is where AI needs to run (The GPU is for serialised pipelines, which is pretty much unusable for heavily stateful AI computations).
The PS4 could potentially compete due to its 8 cores, but writing code that can really utilise 8 cores in parallel is very complex, and oftentimes impossible. You can only have concurrency to some level, so something will always be waiting for another core to finish, losing a lot of that potential advantage.
And then there’s the devil that is procgen. See, most of the seemingly smart AI in shooters aso rely heavily on propper data of the map. That AI don’t see a 3-d landscape when making its decisions. It’s seeing a tree of data points that have been carefully weighted by the developers. ProcGen means you can’t do that. First you need to expand valuable CPU power to generate your actual map, and what you’d need then is a first-level AI that evaluates what was being generated and create a properly weighted pathfinding tree, and you don’t even know what to weight for. Is there going to be a fight on that currently generated terrain section? are people just supposed to stroll around aimless? Or are they on-route to locations that the AI cannot even currently know about with precision, because it’s too far away? And how smart is the creature supposed to use that data, because if every animal suddenly starts running for cover and coordinates with its mates to draw your fire while the others try to outflank you, that’s not really the effect you want either. That’s behavior you’d expect from the Sentinels.
In effect, you’d need to generate several trees, for several scenarios and activities. On top of that the thing has to be dynamic enough so it can insert or remove nodes at will when the player moves and you need to cover the newly generated terrain. And all of this can’t really be paralelised, because the AI doing that has to have the full local terrain available (ergo wait until that is finished), and anything supposed to use the data outputed by this has to wait until that is available, otherwise they’ll just move around aimlessly as they are doing already until that generation has finished.
I won’t go as far as to say it’s impossible. But it’s never been done before, and would be a serious challenge to anybody, even companies like Google, who pretty much own the research field of AI at the moment. As a result, I’m not seeing it happen.
That said, there are minor things where a bit more smarts in the code could help and wouldn’t be too hard to implement. For one, I don’t consider the Idea of MacForaDay to use the data gathered by base sharing to train an AI to build more diverse bases oo far out. It was pretty much the first fought that occured to me when I heard about base sharing back when pathfinder came out. Not neccessarily that they are actually doing it, but that it would be a neat opportunity. Maybe something would come out of it, maybe not. That’s the trouble with trainable AI, you’re never quite sure if it works out. But it’s also reasonably cheap enough to give it a go.
A lot could be improved in space combat, because AI for that is relatively trivial (as compared to AI that has to be terrain aware). But then again, if they do anything about space combat, they should first of all make an actual interface for it…
Creature behavior should be improvable, and I fully expect it to improve a bit, I’m just not sure to what extent. Something simple like having creatures move with more purpose, like herds on migration or packs of predators attacking with some minimum of logic and organisation should be doable, but that also depends on how exactly creature movement and distribution is spawned by the procedural engine. Do they just randomly spawn in a place according to a generated probability map, or is the system smarter and employs probabilistic distribution functions that give flocks of creatures a semi-predictable path over the planets surface? All of that stuff is pretty much unknown, and influences a lot of how much could be done with how much effort.