There is no reliable way to detect AI writing. It probably trains on texts known to be AI, on texts known to be written by humans, then classify the text according to this training.

The problem is that it has a pretty high false positive rate. Maybe it thinks it's AI because there are absolutely no spelling mistakes. Or maybe you're French and you use latin-roots words in English that are considered "too smart" for the average writer.

And the problem is that people run those tools, see "80% chance to be written by AI", and instead of considering that 20% is high enough to consider you don't know, will assume it's definitely written by AI.