[benlivengood]

My guess is that it's pretty simple. There are ML models predicting the ad quality score; when the predicted outcomes for a site/account don't match reality it is a strong signal that something is wrong. Other inference models determine what is likely being violated on the site or by the account. A human has no chance of digging through the model to figure out exactly what causes the flagged outcome. The only solution is trying various changes on the site and seeing what the models and clickers/consumers think of it unless you want to try to pay Google to deepdream your site for you with their inference models; good luck with that.

It's not worth it for Google to fix rare edge cases; fixing them manually may even bias the models to cause problems for a greater number of sites. I am sure the models are trained as well as humanly possible; there are millions of dollars on the table at the fringes. Every employee's time is most valuably spent improving the models as a whole and not chasing down one-off edge-cases unless it's for a very high profile client. I am sure there are whitelists (and blocklists).

Accept AdSense for what it is; a highly-profitable-for-Google way to use ML to turn HTTP into $$$. It will not work for every site, only 99.99xx% of them.

[throwaway45349]

> I am sure there are whitelists (and blocklists).

Off topic, but why is whitelist acceptable but not blacklist?

[y0ssr3n]

I think the preferred term is "allowlist" not "whitelist". I like Allowlist personally because it's immediately obvious what it means esp for less technical folk. It takes time for new terms to become commonplace.

[throwaway45349]

Post hoc rationalisation, but okay. My non-technical friends know exactly what a whitelist/blacklist are but none of them have ever heard these new politically correct terms.