Given that GPT-3.5 (like many LLMs) was trained with a large corpus of scraped internet data, including popular discussion fora, the people on the leaderboard are the ones potentially to blame for ChatGPT’s em-dash habit.