“The LM Arena leaderboard is a flawed method for evaluating AI models because it relies on votes from casual users who skim responses and favor superficial qualities over accuracy.”