The problem described here (LLMs overgenerating their outputs) is fundamental to how they work - they are text generating systems. They generate text! It's not just that they are copying "bad habits" picked up from the internet but they're tuned to output text for large projects.

https://bsky.brid.gy/r/https://bsky.app/profile/did:plc:ttxkeq56hvubnrha24kezy5g/post/3mypxwej25s2m