I find that most of the problem of merging come from lack of randomness. That is both parties think the other should have already acted, so they think it's not their turn to act. Eg. I already slowed down yet the other party hasn't even started merging, so I'll speed up. Then the other car just starts to turn. Bah, I have to slow down. But they noticed I sped up, so they reverse their course, and we're already approaching the end of their lane, so ... poor other car has to now come to a complete halt, and either wait for a bunch of cars to pass, or this lane will have to come to a stop too to let that car in. And so on.

So some kind of consistency would help. If the self-driving car decides to let the other car in front of it, then slow down, wait a bit, then okay, chance missed, go ahead.

How else would it work? This part is classic reinforcement learning. I don't think this policy part is the hard part. (Eg. similar speed merging, and from-stop merging, or stop and let other car in.)

Probably each brand will have a personality. (Based on their timeouts and other parameters.) And humans will adapt quickly. (And naturally the self-driving cars will change over time too.)

Chat requires an amazing amount of implied context. Merging? Not so much. Driving in general? Yes, that does, because there are strange edge cases. (Like what to do when you are lost, or the car breaks down, etc.)