LMArena, a popular benchmark for large language models, has been accused of giving preferential treatment to AIs made by big tech firms, potentially enabling them to game their results.