Yeah, I swear AI writing is worse than it was. Part of my big model release tests has always been asking it to write a few chapters in different styles, and I swear they have slowly gone from really bad, to boring and plain (I remember Gemini 3 being peak of this), to now over-complicated and quite bad. My guess is that they were initially being trained on writing quality but that has taken a backseat since it's not really a profit generator.

I think the signals also show up in different ways to different people.

For example, I don't dislike the way Gemma 4 writes so much.

But I am British and Deepmind's output somehow seems more compatible with the way British people communicate (it's a very multinational team, I suspect that is an influence).

Other people pick up on Gemma's AI writing tells a little more readily than I do. I have a greater tolerance for Gemma (and Muse Glimmer) than ChatGPT or Claude (or Qwen).

They're still heavily RL'd on response formatting/clarity, and against common LLM-isms. Long form prose quality is expensive to RL so I expect this to languish for a while until other low hanging fruit has been picked.