As someone who hates reading AI-written stuff (stylistically, I'm sure that someone could build an LLM that writes well) I have begrudingly accepted that it's here to stay.

However, I think that AI-slop is genuinely harder to read and less information-dense than stuff written by a person, and I catch myself abandoning articles and other written content early on when they smell too much like AI. It's simply less worth it to read them because I'll have to spend some precious time figuring out wtf the text is actually trying to say.

It seems like the flow is "User A writes bullet points and has a model convert them to prose to send to User B. User B then uses a model to convert the text back to bullet points."

I would in every case prefer the bullet points or raw transcript of the AI session to just the final output.

It is harder to read, and it's especially egregious in the human-facing documentation or code comments.

It takes considerable effort, rules, examples, hooks and prompt to keep the popular $model reasonably concise, and yet STILL from time to time it starts vomiting absolute garbage.

At the moment the best recalibration is "shaming" it and showing both the negative (the vomit) and positive (slop replaced with human sentences). Good for a moment (up to 200k context).