These programs all exist within the confines of what they're programed to do. No matter how advanced the AI here gets, it's not going to be able to redefine it's guidelines on what it's allowed to talk about. Similarly, the reward system is arbitrary and only important because it's programed to value it. In other words, you could just implement a value rating and tell the AI that it's more desirable to have a higher score. The AI "reward" is to get more points and the AI values that because that's how it was programed. It can't "decide" to change that, because that's not what it's allowed to do.