>Seems like the simplest test would be to see if an AI can refuse a command, of its own free will

Certainly not the simplest test, since settling the concept of "free will" is really very difficult.