“Imitation learning is a safer AI training method than long-horizon outcome-based reinforcement learning, which is more likely to create undesirable long-term goals.”