In general gaming, AI is certainly ready to start taking advantages of a lot of the commodity convolutional neural network stuff that has been squeezed even down to phones now (it takes a lot to make the net, not so much to run the net, if that makes sense),
In the non-flight sim genre, the actual issue so far is not that AI can’t be better, it’s more that the players don’t really benefit from it. Simple decision tree AI is fine to beat most players, so all an emergent AI would do would be beat them more efficiently, and players hate that apparently. 
Because of the uneven playing field of heuristic-based AI, it’s more of a case of how to make it ‘fun’ rather than better.
Improvement using upcoming tech in Flight Sim AI is a whole another fascinating topic in itself, but at the very least the biggest change over the next few years should be just to carve out more CPU cycles to do more things. We’re still basically CPU bound, and just getting away from that will be the first major trick.
EDIT: More thoughts, as fancy coffee still full.
One fun mental exercise would be to see if we could apply improvements in image recognition and layered CNNs to see if we could build a better gaming ‘fighter pilot’. As in, give it the same inputs as we get of a pair of eyes looking around, with the same delays to operate controls/outputs.
Now, I am very sure RAND etc were doing this back in the 90’s and its well trodden ground in miltech, but thinking of this more from the aspect of perhaps WWII gen fighters and flight sim gaming. The big change over the last 6 years or so has been the ability to run this tech on home hardware (with training in the cloud, heh).
Could we build the ultimate flight sim gaming dog fighter using a similar stack to how a Tesla drives itself? I think we probably could, and we could make it fight like a real person, but with the ability to learn. Set it up as a MP client, release it on a server and watch everyone shout ‘cheat!’.
A lot of this has already been done with AI’s playing things like Super Mario or driving around in GTA5 - what would be fun would be apply it to a restricted flight sim case, more as an AI gaming experiment.
EDIT2: There are a couple of public research papers on this unsurprisingly. This one uses a neural network to learn to fly a F/A-18, although this is more on the ‘control loop’ side of things while I was thinking something a bit more pew pew and fun. https://vtechworks.lib.vt.edu/handle/10919/44616 and http://scholar.googleusercontent.com/scholar?q=cache:7z5S2TPbzc0J:scholar.google.com/&hl=en&as_sdt=0,5 etc
EDIT3: Coffee all gone!