Someone should write this: an blocker that uploads a screenshot of the page, use OCR to find 'Sponsored' and send back coordinates of the box that can then be painted over client side.

Obviously OCR is one of many strategies and as long you kept the code for these adblock strategies open source, it would work on all visual ads.

Train an AI with a billion screenshots of pages with ads marked out. It should be able to handle 90+% of ads easy.