It’s worse than that, false positives are possible but someone generating text should be able to get ai to change some words and formatting to break the watermarking, then ai detectors can tell them how well they did.
I don’t know what the answer but I absolutely know it isn’t this.