4 ms·

> Machines will simply start pretending to be human and sign everything themselves The signature has to be trusted though. Just signing it with a signature tha

> Machines will simply start pretending to be human and sign everything themselves The signature has to be trusted though. Just signing it with a signature that has no meaning to me doesn't change anything apart from perhaps proving it has the same origin as something else. The key function is the humanness or repotuation/trust score of the signature. If I read something signed by my uncle with a key we recently exchanged in person, then that's a high human score. If a blogger I have read and enjoyed 100 posts from posts a 101st post then I trust the content of the 101st to have the same quality (likely because the author is human). Both my uncle and the blogger could be compromised, but it's a start. > nothing prevents a person from taking a ChatGPT output and putting their own, human, signature on it. That would make me lower my trust of that author. I think this is the key: the signatures don't provide much value until we have a good reputation database, perhaps systems of trust networks where my uncles' neighbor is automatically slightly more trusted because I trust my uncle. That would be a useful system. > signing everything does not require any Web3 blockchain bullshit in order to work Indeed.

phoe-krk 4y ago
> That would make me lower my trust of that author. Would it? How would you tell? How can you prove that a piece of text was not written by an artificial intelligence? > (likely because the author is human) I think that's the real problem we're facing right now - humans are no longer the only plausibly intelligent beings on the planet, to the point where the meme-ish line between artificial intelligence and natural stupidity begins to blur. The only way in which you can tell these two apart at the moment is via approximation, just as you've used the word "likely".
- alkonaut 4y ago
  > Would it? How would you tell? How can you prove that a piece of text was not written by an artificial intelligence? You either read enough of them and attribute enough quality that you don't care that it was. This isn't a system to prove whether strangers are human or not. It's a system to keep track of whether their output is human enough for you to bother reading it. If they write 100 things you think feel human, then you can have some certainty that the 101st thing will feel human. That's all. Won't say anything about whether strangers are human. The key here is "likely" and "whether you'll bother". A Turing-like proof of humannes will be possible for the foreseeable future but the caveat is of course a) you don't have time to do that and b) you don't have access to authors even if you did. The social friends-and-their-friends trust networks have a vert quick dropoff too. If you trust your uncle to be human and exchange keys in person (and you'd be good at noticing whether his keys were compromised or he started signing off spam as human for $5), you'll likely never trust his neighbor that you never met as much. But perhaps the total "initial score" of an anonymous signature can be significantly higher if they are in a network you trust. This is directly analogous to normal "trust" where you might rather trust your uncles neighbor to babysit than a total stranger, even though you never met them.
acdha 4y ago
> > nothing prevents a person from taking a ChatGPT output and putting their own, human, signature on it. > That would make me lower my trust of that author. I think this is the key: the signatures don't provide much value until we have a good reputation database, perhaps systems of trust networks where my uncles' neighbor is automatically slightly more trusted because I trust my uncle. That would be a useful system. I don’t disagree but I think this kind of thing will be undercut by people responding to different incentives. For example, maybe your uncle’s friend would never intentionally try to scam you but what if they’re really busy and see some cool AI assistant ad? Or broke and get an ad promising $500/week to send letters from what they’re assured is a legitimate business? Or simply that they use some online service which is hacked in a way which lets the attackers use delegated signing credentials? There are ways to mitigate that kind of thing, of course, but it feels like it’s going to be expensive and unsatisfying.
- alkonaut 4y ago
  Yes, I mean the trust of a friend-of-a-friend will never be as high as your trust in an immediate friend. So the trust networks I think will be small and brittle. Basically: you'll read content from a few people in a very carefully gardened group (Although it can contain things like media outlets etc), and as soon as you try to step outside that, you'll be knee deep in spam and generated nonsense.

phoe-krk 4y ago

> That would make me lower my trust of that author. Would it? How would you tell? How can you prove that a piece of text was not written by an artificial intelligence? > (likely because the author is human) I think that's the real problem we're facing right now - humans are no longer the only plausibly intelligent beings on the planet, to the point where the meme-ish line between artificial intelligence and natural stupidity begins to blur. The only way in which you can tell these two apart at the moment is via approximation, just as you've used the word "likely".

alkonaut 4y ago
> Would it? How would you tell? How can you prove that a piece of text was not written by an artificial intelligence? You either read enough of them and attribute enough quality that you don't care that it was. This isn't a system to prove whether strangers are human or not. It's a system to keep track of whether their output is human enough for you to bother reading it. If they write 100 things you think feel human, then you can have some certainty that the 101st thing will feel human. That's all. Won't say anything about whether strangers are human. The key here is "likely" and "whether you'll bother". A Turing-like proof of humannes will be possible for the foreseeable future but the caveat is of course a) you don't have time to do that and b) you don't have access to authors even if you did. The social friends-and-their-friends trust networks have a vert quick dropoff too. If you trust your uncle to be human and exchange keys in person (and you'd be good at noticing whether his keys were compromised or he started signing off spam as human for $5), you'll likely never trust his neighbor that you never met as much. But perhaps the total "initial score" of an anonymous signature can be significantly higher if they are in a network you trust. This is directly analogous to normal "trust" where you might rather trust your uncles neighbor to babysit than a total stranger, even though you never met them.

acdha 4y ago

> > nothing prevents a person from taking a ChatGPT output and putting their own, human, signature on it. > That would make me lower my trust of that author. I think this is the key: the signatures don't provide much value until we have a good reputation database, perhaps systems of trust networks where my uncles' neighbor is automatically slightly more trusted because I trust my uncle. That would be a useful system. I don’t disagree but I think this kind of thing will be undercut by people responding to different incentives. For example, maybe your uncle’s friend would never intentionally try to scam you but what if they’re really busy and see some cool AI assistant ad? Or broke and get an ad promising $500/week to send letters from what they’re assured is a legitimate business? Or simply that they use some online service which is hacked in a way which lets the attackers use delegated signing credentials? There are ways to mitigate that kind of thing, of course, but it feels like it’s going to be expensive and unsatisfying.

alkonaut 4y ago
Yes, I mean the trust of a friend-of-a-friend will never be as high as your trust in an immediate friend. So the trust networks I think will be small and brittle. Basically: you'll read content from a few people in a very carefully gardened group (Although it can contain things like media outlets etc), and as soon as you try to step outside that, you'll be knee deep in spam and generated nonsense.