That's true, but it's also straightforward to train detectors with massive quantities of human and AI-generated text. So AIs can be trained by the latest detectors, but detectors can also trained by the latest AIs. It will probably come down to which side has more resources to devote to training.
I am concerned about the false positive rate on AI detectors. But if they are being used to flag submissions for further review only then I can go with it. Finding out if a student cheated shouldn't be too hard, just quiz them on what they wrote.
And even before quizzing them, their work can be cross referenced against their other writing. If they've gone from writing with poor understanding of subject matter basics and poor written English to something like an exemplar essay in a matter of weeks, it's pretty clearly they've had some sort of help whether that's AI or not.