It's a matter of decades of tradition. They'd have to change the way this stuff works for this to work properly. If you think about it, 99% of ATC exchanges are entirely routine exchanges of very basic information. Requesting and answering such basic questions as: Who the hell are you, what type of airplane are you flying, what are your intentions, change course to X, altitude to Y, altimeter settings, etc. And that's before we start talking about more complicated approach and departure proceedings.

Essentially all of that can be changed to exchanging information digitally and automatically. In fact most of that information exists digitally. It's not a technical challenge to do that. There are many ways to make this happen. The challenge is entirely institutional because it would represent massive changes to an industry that resists all form of change for decades on end. And that's for good reason: safety. But it also means that a lot of very flaky technology keeps on getting used and that pilots are task saturated with a ridiculous amount of work during safety critical portions of flight. When they get behind on their tasks, dangerous situations happen. When they stop practicing, they lose the ability to stay on top of things. It's that hard. And their pilot license shortly after (currency is not optional).

If you'd design the procedures and communication from scratch using modern tools, you'd end up with a very different system than what we have today. The fun thing indeed is that with modern AI, what might happen instead is that we end up with AI assisted ATCs and pilots delegating the ritualistic exchange of information via VHF radios to AIs. They could just send each other digital messages instead that are then processed automatically and acknowledged manually when needed. But on the off chance of one of the pilots or atcs still being stuck in the nineteen sixties, it's going to be easier to pretend that everyone is and have AIs imitate human actors participating in this.

So, text to speech and speech to text on both ends combined with an AI that has enough contextual information to take simple decisions and control systems escalating to pilots and controllers only when needed to might be more doable than coming up with a new system. I'd say chat gpt 4 would be well up for a "pretend you are an ATC controller" type role playing session.