We just need to give the AI memories from real people and give them a life span short enough that they won't realize it... and then if they go bananas and start causing harm, make another AI to terminate them... I swear I have heard this strategy used elsewhere.