I think the best way around the whole Learning AI debate is this:
From a fundamental standpoint, the game might decipher inputs as all of them going to Popo. And if Nana is able to, she follows Popo. I think in reality we would then have a situation where we see the Nana doing all kinds of grabs because of the amount of chaingrabbing we do on our Wiis. But for my part of playing, I haven't seen any evidence of her doing that on her own. So the computer AI might make a distinction between if Popo is available and how inputs affect Nana and if Popo is not availabe and how those input affect Nana. If Popo is unavailable, we aren't in control of that situation, meaning then that we can't realistically "teach" nana to do anything. And since the that situation is the only time the AI would realistically be in effect, it might all be a moot point.
But in all seriousness, if the AI were "learning", the real question we ought to be asking ourselves is how to best maximize potential advantages of this.