Re:OpenAI has published the text-generating AI it said was t...
by rm_-rf_slash on Thursday November 07, 2019 @02:52PM
(#21476678)
I very much disagree.
Is GPT-2 capable of creating fake news articles? No, but thats a fundamental misunderstanding of propaganda.
Propaganda is a matter of connecting ideas to feelings and creating an ecosystem that reinforces those connections.
For that reason, you need propaganda directors to create the “fake news” or whatever, but where GPT-2 can be very, frighteningly useful, is with bot swarms.
Just a sentence or two from GPT-2 can be convincing enough to be indistinguishable from a human. Have enough of them primed to direct the pro/con conversation online, and people will respond accordingly, making the “conversation” seem real when it’s actually been carefully crafted.
You can already get this today, but that requires armies of people posting on social media. When anybody with access to enough bots on enough social networks can harness this, you can warp any conversation, anywhere, at almost zero marginal cost.
This is something GPT-2 can do right now. Even the smaller models were good enough at producing indistinguishable text. Nobody is immune to propaganda.
Is GPT-2 capable of creating fake news articles? No, but thats a fundamental misunderstanding of propaganda.
Propaganda is a matter of connecting ideas to feelings and creating an ecosystem that reinforces those connections.
For that reason, you need propaganda directors to create the “fake news” or whatever, but where GPT-2 can be very, frighteningly useful, is with bot swarms.
Just a sentence or two from GPT-2 can be convincing enough to be indistinguishable from a human. Have enough of them primed to direct the pro/con conversation online, and people will respond accordingly, making the “conversation” seem real when it’s actually been carefully crafted.
You can already get this today, but that requires armies of people posting on social media. When anybody with access to enough bots on enough social networks can harness this, you can warp any conversation, anywhere, at almost zero marginal cost.
This is something GPT-2 can do right now. Even the smaller models were good enough at producing indistinguishable text. Nobody is immune to propaganda.