Funny thing is correct json is easy to "force" with grammar-based sampling (aka it literally can't output invalid json) + completion prompting (aka start with the correct answer and let it fill in whats left, a feature now depreciated by OpenAI), but LLM UIs/corporate APIs are kinda shit, so no one does that...

A conspiratorial part of me thinks that's on purpose. It encourages burning (read: buying) more tokens to get the right answer, encourages using big models (where smaller, dumber, (gasp) prompt-cached open weights ones could get the job done), and keeps the users dumb. And it fits the Altman narrative of "we're almost at AGI, I just need another trillion to scale up with no other improvements!"