The AIs aren't using emdashes because they're "massively represented in the training data". I don't understand why people think everything in a model output is

The AIs aren't using emdashes because they're "massively represented in the training data". I don't understand why people think everything in a model output is strictly related to its frequency in pretraining. They're emdashing because the style guide for posttraining makes it emdash. Just like the post-training for GPT 3.5 made it speak African English and the post-training for 4o makes it say stuff like "it's giving wild energy when the vibes are on peak" plus a bunch of random emoji.