That's basically how "AI detectors" work, they're just ML models trained to classify human- vs LLM-generated content apart. As we all (hopefully) know, despite provider claims, they don't really work any well.
I think (currently) the problems are more about text, or post processing of other media to hide AI.