1. Training data moved toward dash-rich prose
Root causeWhat changedGPT-3.5 rarely leaned on em dashes this hard. GPT-4o-style output uses them roughly 10 times more often in many everyday drafts.
Why that mattersThe timing lines up with 2024, when AI labs added more digitized, high-quality printed books to training pipelines. Edited books use em dashes heavily, so newer models learned that rhythm.