A longtime standing solution to hard-to-read captchas are easy to read (but hard to process) captchas. E.g. show a picture of an animal or a shape and ask what it is, or even just ask a simple math problem or riddle in writing.

Problem with those are that they need to be constantly updated from a reliable source, or else, once the solution becomes popular, the spammer can bruteforce it in linear time (no matter how high N is, there are only N possible patterns).

This seems to be an attempt at fixing this. Ads are often relatively short lived so by the time the spammer has them bruteforced, it might be out of circulation - and more importantly, there are new ads in. It's an armsrace, and this is a way to pay our troops. Also, it's trivial for advertisers to make many different variations (e.g. one for each sales bulletpoint), so there are many variations in circulation. Since there's often more textual content than the password in the ad, they're not prone to simple OCR, while still easy to comprehend for the user.

Obvious shortcomings are if ads are not so shortlived, and if it's easy to identify and break classes of ads (e.g. if it's yellow and has the IE logo in position X, OCR area Y, done). Also, it's a bit of a dealbreaker if I'm forced to open and visit a website to get the password.