The ethical issues will rather be the moment when AI's are clever enough that they can be teached. What will humans teach them? Personally I believe the best protection from abuse will be to get that learning process in the public (as compared to AI's learning from companies or military for specific purposes). So my own long term target are distributed virtual worlds in which AI's improve by getting passed around. A single computer user might still teach them bad stuff, but I hope some sort of selection based on many people watching and exchanging bots will get the best possible results even though completely preventing abuse won't be possible.
Erasing knowledge and memories of such AI would be analog to killing, and it will become problem, because throwing out data is easier, and writing you enemies to DVD after each game of Quake 10 becouse of moral issues will be problematic after some time.