But employee A can search for a very specific thing, something that machine learning won't be able to do. So it does not need to scale if they want to target a single person.

Employee A has no clue which person to target. Nor does he know which emails to pay attention to. His search tools are primitive. His brain's processing power is limited. OTOH, Google's core business is scalable search.

> Employee A has no clue which person to target. Nor does he know which emails to pay attention to.

Yes, he does: he looks up the people he knows about.

A person reading your mail is a much more immediate, direct danger than your mails being slurped into the big data silo that can slowly and widely ruin society.

> that can slowly and widely ruin society.

There is nothing to say these "big data silos" will run "slowly." And even if they ran slowly, the impact they could have on your life would likely be far larger than the more immediate impact of an employee reading your email.

You don't need deep machine learning to write a simple script to fetch data from your servers.

If Google can read all Gmail users' emails, Fastmail can read all their users' emails.

I think maybe the point is that reading all your emails seems to be Google’s business model, while FastMail is a service provider that charges users for the service? Also, regarding “your data” being more secure with google... is more like Google is good at securing their main asset, which is Google’s data about you.

Google no longer uses Gmail data for targeting. No person or ai is reading your email from your grandmother, florist, or then confirmation that you bought paper towels on Amazon.