Somehow humans are very good at telling apart SEO crap from legitimate content even without understanding the content or the language itself - SEO crap has some common elements such as ads, affiliate links, a certain page layout, etc.

I remember using an open-source ML model (trained on Buzzfeed article titles) to detect YouTube clickbait based on titles and it worked brilliantly, and that was just downloading some code on GitHub and running it as-is. I'm sure the same could be applied to search results and you could achieve much better quality if you actually put some effort into it.

I very much doubt this is some kind of hard problem as opposed to Google just giving up because their business model doesn't actually incentivize good search results.