...do you realize your messages are already in their servers
...but encrypted. With end-to-end encryption (as advertised by many messagning services), the server operators
simply do not have the ability to decrypt those messages.
Even if someone hacks their servers and intercepts the encrypted messages they
can't decrypt them without the private key on the recipient's phone.
However they spin this feature, it can
not work with e2e encryption - the unencrypted messages have to be decrypted on the server to be analysed. The linked papers are just really talking about security protocols that ensure that - pinky promise - can't be accessed by Meta. Just the, er... Meta-written AI software that does the summaries...
and you already blindly trust their claims about encryption, right?
Sure, you can never be 100% certain, but the question is, what is more likely?
(A) A messaging service offering end-to-end encryption is flat-out and maliciously committing fraud and is knowingly using a backdoored system to read your messages - and run a very high risk of being found out via. analysis of the client software.
(B) A messaging service offering not-really-e2e but, hey, shiny AI features!!! suddenly has an "aw, snap!" moment and finds out that the supposedly safe container containing the unencrypted messages isn't has a leak...
Thing is, e2e is the gold standard for secure communication
even through an untrusted channel - that's the whole point. Anything else - where something in the middle has the ability to decrypt the messages - is fundamentally less secure. For the carrier, too, using e2e is the best way to resist demands from authorities.
There's also the question of whether the AI is "learning" from your messages and could leak details to other users...
So one would get the AI summary and then have to read the messages anyway just to make sure the AI hadn't screwed up.
Yup. My main objection .
If you have end to end encryption turned on, how does it summarize?
By turning off end-to-end encryption and letting the server decrypt the messages. In some sort of secure software environment that can't possibly be read by unauthorised persons, honest. So that's alright then /s.
Nope. On the server. If you follow a couple of links from the original article to see the description of "Private Processing".