5 ms·

This is moderately interesting work but the title feels wrong to me. They use LLMs to analyse social media and turn unstructured text ("I always get stuck wait

• anon25783 3y ago
  Would you want it done to you? I sure wouldn't. Too bad it's probably going to be done to me anyway.
  • nl 3y ago
    It pretty much already is right (even if the technology used is different)? A company like Google builds a profile[1] from my search queries. Many other AdTech companies do similar things. > Would you want it done to you? I sure wouldn't. I don't mind this - if I have to see ads I prefer ones relevant to me over ones that are completely irrelevant. [1] https://myadcenter.google.com/controls https://myadcenter.google.com/controls
    • dotancohen 3y ago
      Would you like political ads micro-targeted to you? Gay men get this ad: Only Politician Foo protects gay security in the streets at night! Homophobes get this ad: Then you must vote for Politician Foo, only he opposses gay marriage! Pro-lifers get this ad: Vote for Foo to protect the children from the very first moment! Pro-abortion get this ad: Foo protects women's rights! Advertising should be as general as possible - microtargeting people in a democracy is extraordinarily dangerous.
      • lovepronmostly 3y ago
        wat? targeting markets has been a thing forever. Rim ads in car magazines. Jewelery ads in fashion magazines. I suspect there are different ads in the sports section of a newspaper vs the business section, vs the lifestyle section. I suspect any ads on grinder are different from ads on tinder as well.
      • cornel_io 3y ago
        None of that seems upsetting to me at all. Even if it's advertising, even if it's about stuff that's political, even if it's aimed at causes I dislike.
      • nl 3y ago
        I think misleading advertising is always dangerous but there isn't anything particularly wrong about targeting.
      • upupupandaway 3y ago
        Politician Foo sounds like a major flip-flopper. /s
      • adolph 3y ago
        > Advertising should be as general as possible - microtargeting people in a democracy is extraordinarily dangerous. I don’t understand how that follows. Your examples of microtargeting seem helpful since they speak to how politician Foo appeals in ways relevant to a particular segment. Instead of unknown-likelihood inference of how the politician would operate because of some associations, a voter would be able to make an informed choice based on what is significant to them. The alternative to some level of targeting or relevancing is broad branding exercises like “Coke is it” or “Team Blue protects you—team red wants you dead.”
        • withinboredom 3y ago
          microtargeting in a democracy means that literally every candidate is your candidate. There isn't any way to actually figure out what the candidate's platform is, because it will be tailored for you, specifically ... and it all might be a lie or so misconstrued that you vote for someone who is actually against the things you want.
    • jstanley 3y ago
      > if I have to see ads I prefer ones relevant to me over ones that are completely irrelevant. You don't have to see ads. But if you do, I don't see why you'd prefer the ads to be specifically optimised to trick you into doing something you weren't already going to do. It's not just a question of what things are being advertised, but also how they're framing it.
      • nl 3y ago
        Who said anything about tricking? That seems to be a different issue.
  • hermannj314 3y ago
    One of the examples in the paper was someone saying "I lived in Glendale and was finishing University during the Left Shark incident" and the authors hid the word Glendale and GPT4 still knew what Left Shark incident means and could guess the person's age and location based on where the Super Bowl was that year. The person that wrote the original text clearly knew they were leaking their age and location when they wrote that statement. Why are we flabbergasted that GPT can still infer that info with mild obfuscation? I asked GPT4 to obfuscate the original text and it was smart enough to remove the mention of the Left Shark incident. So maybe using an LLM to sanitize all your posts online is a good defense to an adversarial LLM doxing you. I don't think I see the authors mention doing that at all in their paper (I may have overlooked it)
    • noobermin 3y ago
      > Why are we flabbergasted that GPT can still infer that info with mild obfuscation? May be this was "obvious," but like a lot of research there's a bit of hindsight bias in that sentiment. The issue now is that while you can easily do this yourself in terms of inference, the problem is GPT4 allows this inference to be automated.
• jojobas 3y ago
  > think most people writing on a public site realize their writing is public "Anything you say in public can be overheard and used against you" is somewhat different from "everything you ever said in public will be analyzed and compiled into a case against you" are slightly different shades of grey I'd say.
  • nl 3y ago
    There's nothing in this paper that proposed "compiling into a case against you".
    • jojobas 3y ago
      That's always implied. It's a new capability for governments and influence groups to crack down on dissent of the day. The possible chilling effect is hard to overestimate.
    • withinboredom 3y ago
      "we've canceled your Netflix account, permanently, on behalf of our publishers since it appears that you are accessing [country] content from [actual_country]. We've determined this from public conversations you've had on the internet. Have a nice day."
• SheinhardtWigCo 3y ago
  Automating this in bulk enables societal harms considerably more dangerous than anyone could have reasonably foreseen prior to 2020. If it was discovered that someone had been operating a secret network of CCTV cameras pointed at every street corner on Earth for the last 20 years, would you say "well, you had no reasonable expectation of privacy anyway"?
  • nl 3y ago
    I'm sympathetic to the "doing things in bulk is different to doing it as a once off". But in this case the harm isn't obvious to me. (Also, why 2020? I could understand if you'd said 2016 with Cambridge Analytica but 2020 isn't clear to me).
    • SheinhardtWigCo 3y ago
      Cambridge Analytica collected structured data: friend networks, likes, and interests. At that time, it wasn't plausible to obtain equally powerful insights from unstructured data. That only became a realistic scenario when GPT-3 came out, in 2020.
      • nl 3y ago
        Oh. I assure you plenty of companies were doing NLP based classification in this field much earlier than that. AdSense for example was doing it on pages, and of course Google search strings were a similar example.
  • hackernewds 3y ago
    I don't think your camera analogy is equivalent. It's moreso like actors were filmed with the full knowledge they were, except now the studios are going to make creating deep fakes movies of them available not only to themselves, but to everyone
  • userbinator 3y ago
    If it was discovered that someone had been operating a secret network of CCTV cameras pointed at every street corner on Earth for the last 20 years, would you say "well, you had no reasonable expectation of privacy anyway"? If it really was every, then likely yes. No doubt there would be some interesting footage of important public figures in there too. Not that I'm advocating against privacy in general, but when "I know what you did" is no longer asymmetrical, as in "I know what you did too", it'll certainly change how society thinks about these things.
• cbsmith 3y ago
  > However, the fact that this hand labeling is possible indicates in itself that the issue isn't a pure privacy violation. It's more that automating this in bulk is now possible which was difficult before. Except that it was previously possible, and could be done far more cost effectively, before. Ad targeting systems from fifteen years ago were already doing this with precision in the same neighbourhood.
  • famouswaffles 3y ago
    Ad companies are so valuable because not just any company can reach the scale or access of user interactions enough to perform this kind of inference accurately. Yes you could get precise information but not just from a few sentences said by one user. There's also the quality of information you could previously glean from just unstructured data (noticeably lower)
    • cbsmith 3y ago
      That's not exactly accurate. Ad companies definitely need to train using large data sets, but they could infer with very little information, and in most cases, "unstructured" and "low quality" data was actually quite helpful for inferences about a user. Just a free form text box where someone was prompted for "favourite movies" (with no validation of the content... you didn't even have to have any movie titles in it) could yield surprisingly accurate inferences about income, education, gender, religion, age, marital status, etc. Ironically it was more predictive than inferring from structured data.
• bigbacaloa 3y ago
  [dead]
• swatcoder 3y ago
  > most people writing on a public site realize their writing is public Most people do not think that personal details they didn’t intend to share in their writing can be surreptitiously deduced from things like their word choice and phrasing. It’s true, and people need to learn about and consider it, but no, most people have no idea that it’s really possible or how far it can go. Their intent is to share what they intended to share. Efforts to go beyond that — whether through LLM’s, earlier data mining techniques, or Holmes/Poirot-like observation — are precisely “violating privacy expectations”
  • batch12 3y ago
    While playing with semantic categorization, I also found that it's also what you write about and how you write (seems obvious, but I wanted to confirm). I took a year of hn comments and compared my previous account to my current. It was in the top 20 ranked by similarity (out of something like 60k) just by building a custom 'fingerprint' for each user.
• naasking 3y ago
  > Is this ethical? It isn't obvious to me that it's not - I think most people writing on a public site realize their writing is public, and thew fact we see common warnings about not posting personal information shows that is true. It's not clear it violates privacy expectations either - even if some people don't like it. People are clearly aware of the fact that it's possible to monitor their movements from space, therefore it's not unethical to launch 1,000,000 satellites to continuously monitor all movement on every square inch of the Earth. Somehow I don't think this really follows. The constant mistake I see in these kinds of threads is thinking that "scale" in and of itself has no bearing on ethical arguments. This is obviously false; 1 person pissing in a local river is not a big deal, millions of people doing it is a problem, and genocide is clearly worse than murder.
  • pixl97 3y ago
    Quantity has a quality all its own. Or, I'd rather have a snowball thrown at me than an avalanche.
• gwern 3y ago
  > However, the fact that this hand labeling is possible indicates in itself that the issue isn't a pure privacy violation. They can use it on profiles where that wouldn't've been possible before, and which do not have the obvious tells like "By the way, I live in Melbourne, Australia". Once an 'easy' profile has been labeled, then the model will be trained to learn all the features in that profile which correlate with living in Melbourne, including the ones no human had ever noticed before, and which will be present in other Melbourne profiles. (You could, and should, also remove the obvious tells from the training data after you have constructed the label, to force it to discover subtler features beyond the ones your human labelers used.) This lets you bootstrap your classifier to extract information no human could have before. That would seem to be a 'pure privacy violation' by your definition.