Laws of thermodynamics suggest it will be easier for ad companies to find a way to spam you, than for you to bypass all of their ads. These real-time AI detectors cannot be very cheap to run/train.

If a human can tell what content is an ad (which is also required to be disclosed by the platform anyway), then an AI should be able to detect it fairly easily too.

They have a financial incentive to bypass our LLM ad-blockers but visual recognition is a fairly milquetoast function for LLMs even today. They will attempt to present ads in an unusual way to fool the models; the models will get updated; they will change their tactics again. It's the same cat and mouse game we play today but fought with LLM models instead. I'm confident we win such a fight.