BUTLERIAN JIHAD WHEN?

    • ReallyActuallyFrankenstein@lemmynsfw.com
      link
      fedilink
      English
      arrow-up
      10
      ·
      2 months ago

      The first week after Sora was released, characters were shown doing all sorts of brand-unsafe things. OpenAI spoke with companies after that and created blocked term lists to prevent characters and certain interactions with them.

      That is fairly coarse “limiting.” However, user red-teaming still makes it possible to show copyrighted characters doing objectionable things. That will likely continue once this is released, unless they are doing something different (like having output be re-reviewed by a secondary AI system and actively exclude the output’s description of negative portrayals rather than just user prompt terms).