[angry_octet]

That is underestimating it. DARPA already showed dogfighting that can beat humans in simulation, and they may allow this do execute that.

However, dogfighting isn't very interesting, all combat occurs BVR and in the RF/EW domain.

Pilots are already spending most of their time thinking about epistemic problems: can I see my opponent, can they see me, how do I deceive, am I being deceived, how do I avoid being detected/tracked/targeted/engaged, while possibly trying to do that to everything else. AI is not yet very good at that.

[toxik]

Well, they showed better performance if the AI had full state observation... Not very practical.

[angry_octet]

It's true. The WVR manoeuvring and tactics was the focus of the effort, and DARPA prefers to have short and well defined efforts. But I'm pretty sure the companies are working on techniques for red tactics estimation and behaviour identification.