ChatGPT’s new AI store is struggling to keep a lid on all the AI girlfriends::OpenAI: ‘We also don’t allow GPTs dedicated to fostering romantic companionship’

  • cyd@lemmy.world
    link
    fedilink
    English
    arrow-up
    53
    ·
    10 months ago

    Liability. Imagine an AI girlfriend who slowly earns your affection, then at some point manipulates you into sending bitcoins to a prespecified wallet set up by the model maker. Because models are black boxes, there is no way to verify by direct inspection that an AI hasn’t been trained with an ulterior agenda (the “execute order 66” problem).

    • Kittenstix@lemmy.world
      link
      fedilink
      English
      arrow-up
      5
      arrow-down
      1
      ·
      10 months ago

      Yep, I was having a conversation with a guy that informs policy makers on ai, he had given a whole presentation to a school board meeting I went to a few nights ago.

      He said that’s his highest recommendation when it comes to what should be done on the lawmaker side, pass bills that push for opening up those black boxes so we can ensure transparency.

      • cyd@lemmy.world
        link
        fedilink
        English
        arrow-up
        8
        ·
        10 months ago

        Problem is, there isn’t a way to open up the black boxes. It’s the AI explainability problem. Even if you have the model weights, you can’t predict what they will do without running the model, and you can’t definitively verify that the model was trained as the model maker claimed.

        • Kittenstix@lemmy.world
          link
          fedilink
          English
          arrow-up
          1
          ·
          10 months ago

          I see, my knowledge is surface deep so I admit this is new information to me.

          Is there no way to ensure LLMs are safe for like kids to use as a tool for education? Or is it just inherently going to come with some risk of exploitation and we just have to do our best to educate students of that danger?

    • dirthawker0@lemmy.world
      link
      fedilink
      English
      arrow-up
      3
      ·
      10 months ago

      Some guy in the UK was allegedly convinced by his chatbot girlfriend to assassinate Queen Elizabeth. He just got sentenced a few months ago. Of course he’s been determined to be psychotic, but I could imagine people who would qualify as sane getting too deep and reading too much into what an AI is saying.