Design by AI? No, implementation by AI!

[François Luc Denhez-Teuton]

I don't think we should worry yet about AIs designing games, at least until we figure out how to make A actually I, if you get what I mean.

If you don't, I'll explain. I see intelligence as at least two things: 1) recognition/identification/model-generation (AKA Input, or "Acquiring a working model of the world"), and 2) Problem-solving (AKA Output, or "Organizing that which I have power over to achieve my goals"). I want to say that there's at least a third thing, which is "figure out what my goals are", but I don't know whether that's actually intelligence or not. Certainly there's intelligence involved, but it's more that you have one central goal, which is immutable, and then all other goals are derived from or compared to that goal.

Anyway, the question of "Design by AI" is thus a question of 2 things: what the goal of the AI is, and its problem-solving capacity. If I were an AI builder, I would make a problem-solving program first, which receives as input a user-defined problem, and outputs a program which solves the problem if it can be solved. If I input the problem that would be solved by a game, obviously it should output a game. But the game would be designed by me, at least indirectly. The program would only solve the problem that I set it.

Then there's the necessity for the AI program to know how to achieve certain goals, like "make the player feel relieved" or "frustrate the player" or so on.

My point is that until we make an artificial person that can set its own goals, you're never going to have design by AI, because the goal will always be specified by the human programmer! Or it's not going to have a goal, at which point it becomes indistinguishable from background noise, which is not what we want, now, do we? (From AI, at least.) Put another way, I think that the problem of AIs designing games requires us to know how to make complete AIs, which I believe makes this problem AI-complete.