Why does AI overuse the em dash?

It's a token problem.

", and" = 3 tokens — = 1 token

AI models are rewarded for brevity and clarity. The em dash does both in one character.

Then fine-tuning reinforces it: humans rate concise responses higher, so AI learns that em dashes = good writing.

The model didn't develop a stylistic preference. It optimized its way into one.