I'm willing to bet it's because model providers have gone all-in on programming and reasoning ability when training new models. It's where the money is, if nothing else, and the writing quality will have suffered as a knock on effect.
It was the newness and surprise that the quality was as good as it was, just like images from Stable Diffusion, etc. Now that we are used to it, we notice the uncanny valley speak.