Let me first start by saying I write software for a living so this idea intrigues me. In fact, I think it would be a very cool comp sci grad. program to write. Okay, so anyways...
Think Chess and Deep Blue... it will happen just like it happened in Chess. IMHO, if you play an AI bot of the future, heads up say for 100k hands he will win more playing mathematically perfect than a human who can make big mistakes. He will be programmed to make mathematically perfect play, which I think will be a huge advantage. I think you can program an AI to play just as many random hands as live players to throw them off. I think you could even give the AI the ability to change up its play through out a session. (You know laggy for a while then maybe nitty for a while based on certain conditions, etc.)
I think the difference between chess and poker is that there are so many MORE variables in poker and there is more chance that it will be a more complicated AI to write. I think it can be done. Can the AI possible have a "red flag" warning that a player MIGHT be on tilt so as to raise the range of hands the tilting player might play? Why not? AI could determine, the very hand before the player to my right got sucked out on and lost a big hand.
I think the only real way to always beat these AIs would be to play very unpredictably as to throw off and tracking that the AI does (such as figure your range as 20% or that you tend to more often than not bet the river for value or xyz other contrived ideas). But even in these situations the AI could still possibly pick up on a better pattern or time it took to bet or SOMETHING to give it an edge... I think it can be done and may be done by some brilliant mind at MIT or something in the future.