The absolute worst case scenario would be somehow tricking the AI or taking control from it so jets start attacking/kamikaze-ing their own side en masse. I suppose this is theoretically already a risk for some UAVs and perhaps even some advanced human-piloted jets, but a hijacked autonomous fleet would be a wild sight to see.

No doubt some organizations are researching how to perform and defend against things like that. The future of asymmetric warfare in particular could get really crazy.

If I’m approaching an AI equipped force with my own AI force I’m driven to think about the ‘meta game’.

In this context I would sacrifice multiple aircraft to ‘taint my the enemies training data with solutions which are a disadvantage in a future fight scenario.

Good point, but the meta will surely be played by the other side, too. They'll expect that their enemies will want to try to "obfuscate" their true operations to confuse things that have been trained to recognize and predict certain actions.

That's why this probably won't actually be in use for many decades. You need something a bit closer to AGI than what exists now. Until then, you'll probably have limited use cases, maybe where movement and obstacle avoidance are autonomous but everything else is programmed or remotely controlled.

You could imagine small, agile flying assassination drones which can navigate cities and houses to kill targets with short-range projectiles. Very Black Mirror-esque, but far more ethical than the current approach of Call of Duty 4-style blowing up a big patch of land from the clouds with a Hellfire missile, including all the civilians in proximity.

AI doesn't learn in it's production form. That's training. You'll get the same deterministic response pout of every fighter, even if there is So.e level of sensitive dependence on initial conditions.

Human pilots do mot suffer this problem. A human can arbitrarily change things up for better or worse. Going the AI route is basically going the T-34 route. Suddenly quantity (via reproducibility) takes on a quality all it's own. My issue with that philosophically is does going down that road actually solve any pressing problem?

Which to me it rings a bit hollow and empty unless there's so.e other task we'd prefer people that would be pilots be doing.

General AI probably would learn in real time by most definitions, but setting that aside for a moment I was more referring to gathering production data then using it later. You can gather then re-train, isn't that what Google is doing with our voice data for example?