AI is penalized for any and all emitted errors through RL, so this kind of carefulness is actually a strong sign of AI.