Would you want it done to you? I sure wouldn't.
Too bad it's probably going to be done to me anyway.
Too bad it's probably going to be done to me anyway.
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)
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.
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.
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.
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.”
The voter could not make an informed choice in the case of targeted advertising because there would be no information. There is no signal, everything is noise catered to a particular user. Yuval Noah goes in depth with this were he imagines a world with human like AI agents that befriend us for the express purpose of politically manipulating us at a later date.
https://web.archive.org/web/20230428190758/https://www.econo...
1. Let's say there are two parties: pink and purple.
2. You absolutely love Pink Party because they want to paint the roads pink and force everyone to drive pink cars.
3. I'm running for Purple Party, which says everyone should be allowed to drive whatever color car they want, but we also want pink roads.
4. Gerrymandering is illegal, but our team determined that we can sway some areas. You live in one of those areas.
5. My team writes up an ad saying something like, "Vote Purple and support pink roads and pink cars!" We target all the Pink People living in the aforementioned areas.
6. Pink party runs a very similar ad.
7. You get confused, wasn't there one of them that didn't support pink cars?
8. You google the parties, and visit our website to read our platform. We see that you were in the ad segment, so we hide our support of any color car and change it to say that we're for pink cars.
9. You give up, and just decide that you like my beautiful face -- but more likely you just didn't see the difference, so you vote Purple Party at the election.
10. I get elected and we get pink roads. However, the cars are quite multi-colored.
Hopefully, you can see how this is dangerous? It's quite different than without targeting (where everyone sees the same message). The ads I saw told me what I wanted to see, while the ads you saw were what you wanted to see and confused you enough to vote against your own party.
1. Google and party website: this can already be done and it is easily detected if a person is interested in a topic since other sources will broadcast the party position on a topic.
2. Party-centricity: depending on your locale, people vote for a person not a party. Party affiliation is often a proxy for how an elected person would behave but there are typically differences between party and people. Those differences are critical for voters who do not identify with a party.
In the original hypothetical, candidates who differed from their party were better able to articulate that difference without broadcasting messages that alienated party stalwarts. I see this as a plus for democracy since people elect politicians whose behavior is more in line with voters’ interests.
On the other hand, I can see from your hypothetical that segmenting messaging campaigns allows liars to hide lying a bit better, which is a clear negative.
I suspect any ads on grinder are different from ads on tinder as well.
/s
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.