Private AI?
WIRED
How to Use AI With Your Privacy Intact
Your conversations with AI chatbots are both highly personal and deeply vulnerable to surveillance. Here’s how you can protect yourself.
Very confused, is this all BS?
“Confer, the private AI chatbot launched by Signal creator Moxie Marlinspike, uses that Nvidia-based TEE setup— along with an authentication system…”
“Meta, too, uses a TEE to offer a more private version of the Meta AI chatbot included in WhatsApp—a privacy system that Confer’s Marlinspike has said he is helping Meta to develop.”
This feels too good to be true and I am surprised this article is the first i am hearing about the relationship between Marlinspike and Meta for this product.
I’m no expert so take this with a grain of salt.
But basic logic says if the data (your prompt) gets processed by anything then it needs to be able to read it. Therefore its not private by default.
Sure, they can encrypt it (where? At rest? End to end?) and not retain history etc, and that can make it more private, but it won’t be private.
That’s like Proton’s Lumo which is “private” in the same sense, but the requests are being sent to a server to process so that data is read by something somewhere. I use Lumo BTW, but never put personal info into it etc.
Then no matter how private it is, if Meta starts using it it’s immediately spyware.
If you want true private you’ll need to run your own local models.
The TEE mentioned in this article is essentially hardware level identity/encryption similar to the boot chain of a phone or game console, so if you trust nvidia then it sounds like it would work. If you trust nvidia. And all of the other implementation details.
There is actually Homomorphic encryption where you can operate on encrypted data, but I don’t imagine it’s used much in machine learning since it takes 100-10,000x longer than the plaintext operation.
It’s two different problems :
technically speaking TEE and homomorphic encryption have proven to work. You can read public research and results from e.g. NIST or ask a researcher in the domain you trust but basically unless you find a way to break it (which is not impossible theoretically, just nobody has ever done it), it is indeed private.
in terms of business model, Meta is an advertising company which only happens to rely on tech to sell more ads. The more information they have about someone, the more expensive they can sell ads to their true clients, advertisers. So… one can safely assume that everything AROUND the encryption itself will be used to align with that very profitable business model.