From what I have read and seen:
The data about the non-native detection being much higher than native is true.
This is likely because of a much higher usage of software such as Grammarly, pro writing aid, and other grammar and spell checkers.
Why is this relevant? Because these are powered by AI, and have been for a long time.
The data originality.ai put out claiming their accuracy was 97.4% accuracy.
https://originality.ai/blog/ai-content-detection-accuracy
From personal testing, I can say this is hardly accurate (lol)
Although it is VERY good at spotting unedited generations from chatgpt. I'm not sure how helpful this is, because ChatGPT's voice is very strong and obvious to me (and probably a lot of others) that have used it.