I wouldn't hate this, even if it's imperfect - the most common false positives would be brand-pushing content I don't want even if it's not actually paid advertising.

You wouldn't need machine learning, just allow the image to be seen by a hundred users and see how many mark it as an ad.

Well, that's how you build a training set.

But you still have to identify features and use an algorithm that basically goes "users identify images with features A, B, and C as ads, and they identify images with features X, Y, and Z as not ads, so put positive weights on A, B, and C and negative weights on X, Y, and Z, and block anything with a negative enough score".