LMArena presents itself as a leader in AI evaluation, but its methodology is fundamentally flawed. Researchers and companies often treat the leaderboard as a benchmark, yet it rewards superficial metrics over genuine accuracy. The system relies on internet users who make quick judgments based on aesthetic appeal rather than substance. For instance, longer, well-formatted responses with emojis often win votes, even when they contain significant inaccuracies. A recent analysis found that 52% of user votes on LMArena were incorrect, highlighting a serious disconnect between what users perceive as correct and what actually is. The openness of LMArena allows for easy manipulation. With no quality control or accountability for participants, the leaderboard becomes a gamified platform where flashy presentations overshadow factual correctness. Examples illustrate this problem: a response that inaccurately quoted "The Wizard of Oz" won because it seemed more confident, while a mathematically correct answer about cake pan sizes lost for the same reason. LMArena’s leaders acknowledge these issues but continue to promote the platform as a legitimate evaluation tool. They attempt to implement corrective measures but are essentially trying to fix a broken system. The consequences are significant. As the AI industry increasingly aligns itself with LMArena’s flawed metrics, models become optimized for style rather than accuracy. This misalignment leads to the proliferation of unreliable AI outputs, undermining the integrity of the field. While companies feel pressured to conform to these standards for market competitiveness, there’s a growing call for a return to principles that prioritize accuracy and utility. Some labs resist the temptation to chase the leaderboard's allure, focusing instead on producing quality models that users appreciate for their reliability. The choice between short-term gains and long-term integrity is stark, and the industry must confront this reality.