That's an interesting point, but I'd contend there is a difference between scanning for known virus patterns vs. feeding your email into ML algorithms to do God knows what with.

If someone comes to Google, asking for the content of someone's email, is Google technically unable to provide that information for past emails?

Because I am aware of no reason to think that Google stores my gmail with zero access. I don't know for a fact that ProtonMail discards this information at the earliest opportunity nor do I know for a fact that they don't try to aggregate it to learn about you (or even people in general), but that is what I interpreted the pitch as.

But, look, of course if they get a subpoena they will have to start scanning your email if they are technically able to collect it. That's just a wiretap, and little would prevent the author and operator of the server software from doing whatever they want... and they're clear that if you aren't sending email between two compatible accounts that there is no E2EE.

We can talk about how they should have been clearer about the need to use Tor to avoid IP logging (even if they don't do it, someone between you and ProtonMail certainly could). That's a good idea. But they are actually very clear that E2EE with your email is not what you should expect in general. And I don't think they have much incentive to scan my email from unencrypted sources to do anything nefarious, but I don't think anyone has any ability to prove they do or don't at present.