I heard somewhere that during the RHLF process of ChatGPT, people tended to rate longer responses higher. It might be something similar going on here.