"You're honor, it wasn't me, it just deep-learned it that way."

If a program uses AI to come up with strategies, how do you make sure it doesn't learn to spoof? Coming up with the goal-formulation of "don't spoof" seems like it would require the program to introspect about its own intent. Dunno, I would like to hear what kind of metrics one would use to make the learning algorithm reject spoofing strategies. I guess it would be something similar to whatever heuristics the SEC uses to flag potential spoofing, but they get a lot of false positives and almost never convict someone on metrics.