Hiding Meaning from LLMs, the Obvious Approach

Hey guys! So last time we talked about the fact that LLM providers can see all your messages. They can basically know what you’re talking about, build a profile on you, and know you better than you know yourself.

What can we do about it? That’s a good question.

The most obvious solution I could think of is basically hiding the meaning of your queries. So you need some kind of — maybe we can call it semantic homomorphism. You take your actual query, your text, your question, transform it into something completely different, send it to the LLM, the LLM processes it and answers it, and then it comes back to you and you should be able to transform it back into your own semantic space, so that it’s still valuable to you.

For example, the simplest form of it is anonymization. If you send a mortgage agreement, a mortgage contract, to the LLM, you could hide all the details: your name, the builder’s name, the location, the dates. And then just keep all the context around it and still be able to ask the LLM questions and get useful answers.

You don’t even have to hide it — you can replace it with some kind of fake data, and then it won’t be able to attribute it to you directly. Well, the provider will still be able to attribute it to you, because they know that you asked it.

But you could apply different transformations. For example, you could change the meaning of the message: “my friend asked me to analyze this mortgage agreement, could you please help me provide answers to these questions, or make sure that this happens.” That would be one of the ways.

That’s the most obvious one. Another, maybe more complicated way is actually transforming your query — your question, your mortgage document — into something completely else. Like a fiction book, or a novel that you’re writing, but you’d like it to be as realistic as possible, and make the LLM think that that’s what’s happening.

So those are just the most obvious examples I could come up with. I would love to hear if there are actually solutions already on the market for solving this kind of problem, because it seems like this is something that we desperately need.

That’s all for today. Thank you guys, and enjoy your day.