Re:GPT-2 is not as dangerous as OpenAI thought it might be
by tylerl on Tuesday September 10, 2019 @11:45AM
(#20929452)
My understanding of the danger was a bit more practical and less abstract. If you use your "best" classifier in a GAN to generate malicious content, then regardless of whether or not it's any good, it will by definition be able to fool your spam classifier. So fighting machine generated garbage has to be done by humans again, which scales much more poorly.
[ Parent
| on HN ]
Re:GPT-2 is not as dangerous as OpenAI thought it might be
by Bartweiss on Tuesday September 10, 2019 @12:15PM
(#20929825)
That's a really good point, and it never even occurred to me. Thank you!
Something I realize I don't know: how much spam detection is done via text analysis? I get the sense that most websites try to block traffic patterns or apply CAPTCHAs, and email spam filters are heavily based on keywords and senders. It seems like this would cause problems, but it's not immediately obvious to me where they'd crop up.