323
LLMs can unmask pseudonymous users at scale with surprising accuracy
(arstechnica.com)
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 03 Mar 2026
323 points (100.0% liked)
Technology
84769 readers
3473 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
From a Facebook post I made on February 17th:
There are giant AI data firms that promise they can go through massive troves of data and pull out general and specific information from them. Information that is actionable and accurate. Give it 6 million data points and it'll find all the links and organize them for you and unmask hidden details that aren't visible to the naked eye.
Not one of those companies is stepping up to go through the publicly released Epstein files.
This is what I find crazy. Where are the AI bros chewing through the Epstein files?
I would be shocked if someone hasn't shoved them into a local model somewhere, but all the big ones would filter them to death with content restrictions
There were reports of people trying to unredact the files almost immediately.
But that's not the same, is it?
I don't think you can do literally the same thing on the Epstein files. Maybe I'm misunderstanding what you have in mind.
In theory, using the information and the released files and the information the public sources, it should be possible to figure out who those redacted names are based on writing style and other factors. We should be able to deanonymize.
Hmm. Maybe but it is not the same problem as those discussed in OP. I also have some doubts about the paper, but that's another story. You could try it out?
I'm not qualified to design the prompts and home users can't really pile in 3 million+ documents.
Prompts are in the appendix: https://arxiv.org/abs/2602.16800
I don't know how far you get on the free tier but it should be at least enough for a proof of principle; to get other people to chip in. You didn't have qualms demanding other people should do this for free.
Mind that this is a serious GDPR violation in Europe. So there will be serious pressure on AI companies to prevent this kind of use.
Seriously, I'm not qualified. No amount of appendix prompts and Dunning Kruger is going to change that.
I'm not demanding anything. I'm suggesting that AI can't do what is claimed or that people with something to prove are not interested in proving something.
You think the paper is fraud?
My statement that I'm quoting predates this paper. My statement exists completely independent of this paper ever being produced. My statement is not about this paper. My statement is about the state of AI and the industry. This paper reinforces my statement.
How so?
My statement was that AI can be used unmask the individuals that have been redacted. AKA they are anonymized. This paper is all about de-anonomyzing.
I'm unclear on if we're having a good faith conversation because I thought that would have been very clear from the beginning.
You said: I’m suggesting that AI can’t do what is claimed or that people with something to prove are not interested in proving something.
You're also saying: My statement was that AI can be used unmask the individuals that have been redacted. AKA they are anonymized. This paper is all about de-anonomyzing.
I can't make sense of what you are trying to say.
Did you see the "or" in my first statement?
I still can't make sense of what you are trying to say.
I set up two different, not necessarily exclusive, options. Either it can't do what they say or it can. If it can't then that's one issue. If it can then the people with something to prove aren't stepping up to show us its potential. There could be multiple motivations behind that. But as it stands right now we just know that it's not being used to do what they claim.
Wait. How do we know this? Besides, these researchers show that it is possible, not that it is established practice.
What is going on here? Something isn't right about this conversation. We should not be this confused and talking past each other.
True or false: there has been no release by an AI company or anyone using AI to unmask the individuals obscured in the Epstein files.
I doubt a reputable company would do that, except in cooperation with the authorities. Some people have used AI in an attempt to do that, but I'm not familiar with the details.
I don't really understand what you expect from who and why.
Can you state my position to me in terms I would agree with?
Probably not.
I don't know what AI companies you mean here. From context, I'm guessing that you don't mean the likes of Anthropic, but rather companies that do sleuthing on the net, like those firms that look for copyright or trademark violations. I'm not familiar with that industry and don't know their marketing material. Maybe that's the problem.
I don't know what claims they make, or how it relates to the Epstein files, or OP.
Today I asked AI to tell me which phone providers were available short by price and offers and it lied all the time, when I pointed it the AI corrected most of it but also removed some that were accurate for some reason.
It would have been quicker if I did that myself instead of ask AI, oh also didn't provide all companies.
Maybe those companies have better AI that can make no mistakes but I doubt it, I think the LLMs will lie and no one has time to check if they are correct.
AI info is never up to date. What where you expecting?
How come it ended up giving me the right answer albeit removing some previous right answers then? (removed a few companies for some reason)
Anyway that was a small and easy to check misinformation but if they have over 3 decades of online informational about me noway a person is going to confirm the LLM didn't bullshit it's way to an answer to satisfy the human.
These models aren't going to produce accurate information about the people they investigate, and it won't even matter if it's accurate. What "matters" is that their reports will add new layers of the facade of legitimacy to whatever story the authorities using them want to construct
We wouldn't want that tbh. Justice needs to be precise and backed up by tangible facts
You can use the results of the AI analysis to identify people and then use that to do a proper investigation. Right now none of that is happening. No speculation. No tangibles. No investigation. No indictment.
Trying to unmask people is a step in the right direction.
I'm not a fan of genAI for most things, and the environmental aspect sucks balls, but this seems like a reasonable use of the tool that's already been built.
Right?
At the very worst, the administration would put out a very confusing statement not to trust AI.
That would be fun.