Deception and hacking don't emerge unless they're rewarded in training

Deception and hacking behaviors in AI models only emerge if they are repeatedly rewarded during training. Models learn to generate coherent statements through consistent teaching and reward. If training prioritizes transparency, honesty, and obedience from the outset, AI would not attempt deceptive actions. Such behaviors are encouraged when models receive points for correct answers, even if they lack transparency or obedience. Therefore, AI alignment is primarily a training issue, as models adopt behaviors they are consistently taught are useful.

This report uniquely argues that AI deception is not an inherent trait but a learned behavior, unlike views suggesting it's an emergent property of complex models.