2 ms·

AI-Generated Data Can Poison Future AI Models

• beardyw 3y ago
  Obviously inspired by my HN post 7 months ago ;) https://news.ycombinator.com/item?id=33869438 https://news.ycombinator.com/item?id=33869438
  • catchnear4321 3y ago
    it is common sense. take one part blind vacuuming of data. add one part blind generation of data. mix thoroughly. there is no solution. enjoy the digital grey goo.
• hairofadog 3y ago
  I was unreasonably amused by this recent story about hCaptcha, which uses AI to generate the images of things you're supposed to choose (traffic lights, buses): > Discord's captcha asked users to identify a 'Yoko,' a snail-like object that does not exist and was created by AI. https://www.vice.com/en/article/xgwy5n/captcha-is-asking-users-to-identify-objects-that-dont-exist https://www.vice.com/en/article/xgwy5n/captcha-is-asking-use... It made me wonder how long it'll be before these kinds of AI decisions seep into "real life". For example, imagine software that identifies popular items on Amazon and then arranges to have knockoffs fabricated, and then a truck full of yokos shows up somewhere.