Earlier quoted context omitted.
I don't think so. Machine learning is generally poorly suited to dealing with an intelligent adaptive adversary capable of an unpredictable universe of inputs.
Text classification might be very good at identifying ad text. There's a popular fake news dataset where models hit 85% accuracy https://link.springer.com/chapter/10.1007/978-3-030-68787-8_... Ads are highly tuned for click through rates. Even if the model isn't 100% accurate, it would force advertisers to use less effective ad text to avoid filters.
That seems counterproductive to me. I mean, ads would still be annoying. And that’s why we block them.