2 ms·
Personalization of the fake text to precisely match what the reader is most susceptible to is probably a bigger problem, especially if the bad actor is able to
by onion2k 8y ago
Personalization of the fake text to precisely match what the reader is most susceptible to is probably a bigger problem, especially if the bad actor is able to target small groups (say, politicians?). An AI writer that could write what someone will believe, in the style they're most open to, using their personal information and a history of exactly what they've already seen, would be very hard to resist.
Couple that with scale and it'd be game over for distributing written information across the internet. No one would be able to believe anything they see online any more.
Although, weirdly, that actually sounds like a decent use case for a blockchain.
- thom 8y agoOf course, any AI that is sufficiently advanced to sway both public and personal opinion will probably also be able to mount a 51% attack against whatever blockchain we expect to refute its lies.
- Invictus0 8y agoNot so: 1) because the abilities of AI don't scale in that manner and 2) because, unlike in a true decentralized blockchain, there are centralized trust sources that you can use to verify the content. This is really a better use case for keybase rather than blockchain.
- joe_the_user 8y agoWell, most "high-value" groups like politicians, journalists, billionaires and such are targets right now in the sense that intelligence agencies and private opportunists have their information and trying to use text to influence them. The AI we're talking about isn't as good as human and so it's not going produce things that even as well tuned as a people currently do - since the method involves just emulating normal text, the AI is, at best, going to become nearly as "good as average". But it's reasonable to say this could do a bit of damage to "moderate value targets". Given that you already some portion of retirees today "infected" with fake-news obsessions. Not only would have personalized spam/social-engineer but you could train the AI further on what worked once you even had a lowish success-rate. All that said, it seems like the OpenAI text generator would not be such a customized social-engineering-constructor. Rather, such a thing would have to be trained by the malicious actors themselves, who have their own data about what works. So the now-always-in-the-background question, OpenAI's shyness to release code justified, seems like still a no.