As other comment said, coin flips aren’t random. And it is really hard to find something truly random. For example computer random generators are based on current time, temperature,… which is chaotic enough to give us seemingly random numbers.

If you would have truly random generator, I think, it would be really bad idea to use it. Let’s assume, you would listen to it (otherwise the AI could predict you wouldn’t). Because now you have 50% chance the AI is correct, you would in average get 0.5 * 1000000 + 0.5 * 0 = 500000 if choosing only one box and 0.5 * 1001000 + 0.5 * 1000 = 501000 if choosing both. Each could occure with 50% probability, giving 500500 on average. Which is for one boxes (and their main argument of average case) just over half of their expected outcome. And for two boxers, choosing both is always better then choosing at random, which is better than chossing one.

This case with random generator is similar to the case Minute Physics made about someone else choosing for you.

So, overall, it is interesting strategy, but I think it is worse than choosing deterministically.

PS: I think this strategy would make the situation worse for everyone, because you undermine the trust in the AI making correct choices. It’s measured 99.99% correctness would fall towards 50%, based on how many people would do this. And trusting in the AI with 99.99% prediction rate is reasonable for one boxers. But the lower the measured probability falls, more people would switch to two boxes, which would worsen their outcome, even though AI could be almost flawless with people choosing on their own. So imagine going in as a first person, doing this, choosing two boxes based on coin, and winning big. The next person would probably choose 2 too and lose.