Spearphishing is a huge source of compromise at the moment; antiphishing filters might, in that view, be considered a security feature rather than a security fault.

On the other hand, I have, frankly, never understood these privacy arguments. Is it a privacy violation if someone checks a checksum of my incoming mail against a blacklist? What if they compute a hash of my mail to check the DKIM signature? And if those are OK, why is an ML model more of a problem?

No, the privacy concern is about the potential for abuse if there is a known place in the world where a very detailed account of all my online (and some offline) activities, contacts, communication and personal interests is stored.

It obviously arouses the interest and desires of criminals, governments, employers, politicians, landlords, insurance companies, creditors, marketers, ex partners, extremists and bigots, journalists, potential mates, researchers, etc.

Some of these groups are overlapping and some of the concerns might only arise in the distant future. Governments change. Ownership of companies changes. Personal circumstances and opinions change, but you can't take anything back once its out.

So even if someone hacks into the spam filters, they won't be able to reconstruct content from your account. They _might_ have insight into things like word usage, but they won't have reconstructed sentences or the like, a priori.

You can build systems like this without having long-term storage of the content, which protects against data issues for one-time leaks.

An email with the order went to my @gmail. The next day in my Youtube videos the ads where for stop smoking patches.

They read email.

Sometimes hilariously. A few years ago a friend emailed me a barbershop quartet video, and I spent the next couple of weeks getting hairdresser ads ...

And, if it's being interpreted by a machine, does that really count as reading?