Without the pain-like signal active, the two larger Qwen AI models chose harmful options in only 0% to 4% of their first decisions.