this seems like what should be a one week data science project at YouTube. kind of embarrassing
My guess is that YouTube could easily build the exact same tool. The problem is when you roll it out across your entire site, the spammers figure out the holes in it and start abusing it.
I'm guessing the same thing would happen with this tool. If some large percentage of channels started using it, spammers would find the holes in it.
Sure, hackers will always find loop holes. But when that happens we flag that version as vulnerable and release a patch to fix the vulnerability. This is the exact same technique Google could apply to the YouTube spam. They just don't want to spend the money or time to do it.
Typically, when you fix a vulnerability, things are strictly better. Attackers can no longer do X bad action, but all legitimate users can still do everything they wanted.
Spam fighting is different. If you make your spam classifier broader and broader, it will have more and more false positives as well, and legitimate comments will get deleted too. Without AGI, or at least very good language parsing, it really will be a case of tuning between "more false positives, less spam" and "fewer false positives, more spam".
There's also vastly more spam than there are security vulnerabilities since there are hundreds of thousands (millions?) of people intently creating spam for profit, while bugs are mostly accidental, and exploitable ones relatively rare.
P.S.: Even the deleted comments are deleted silently.
They basically need to do the exact opposite of whatever they are currently doing.
Imagine actually referring to yourself as a Googler.