Valve quietly not publishing games that contain AI generated content if the submitters can’t prove they own the rights to the assets the AI was trained on

  • mack123@kbin.social
    link
    fedilink
    arrow-up
    2
    ·
    1 year ago

    I have to agree here. Generative AI has so much potential for games. Especially RPG style games for believable NPC characters. But the rights environment is very murky.

    I expect it to be resolved relatively soon though. a combination of generally trained AI with subject specific training should do the trick. In the same way we would train a helpdesk bot on company specific information.

    The remaining question though is what of the original broad dataset the source model was trained on. There things are less clear.

    • Ronno@kbin.social
      link
      fedilink
      arrow-up
      3
      ·
      1 year ago

      I think that it is quite feasible to do though. Take for example Lord of the Rings. If the game dev has the rights to Lord of the Rings and its books, then it can be completely fine to write prompts for NPC text as: Produce a response to question X as if the NPC is living in Mordor, with his background as a blacksmith etc… AI can then generate that text under the IP of Lord of the Rings just fine. And it will always be the right tone of voice. Same can be done using dynamic events etc.

      • tal@kbin.social
        link
        fedilink
        arrow-up
        4
        ·
        1 year ago

        KoboldAI and similar do that today, but it’ll soak up all the capacity on a computer and then some just for the text generation. Needs to be more efficient than it is today if one’s going to be generating text on the player’s computer.

        If you mean the studio using it to generate static text, then sure.

        • JackGreenEarth@kbin.social
          link
          fedilink
          arrow-up
          4
          ·
          1 year ago

          @tal Or it could make an API call to a server, the way ChatGPT does today. Unfortunately, that will mean the player has to be online to use the text generation, but the tech of it isn’t what we’re discussing anyway. We’re talking about the ethics of it, not the means.

          It’s like we’re talking about whether robbing a bank is OK or not, and then someone goes and talks about how hard it is to rob a bank. It’s a non sequiter, it’s not what we’re talking about.

          @birlocke_ @lengsel @mack123 @Ronno

          • mack123@kbin.social
            link
            fedilink
            arrow-up
            3
            ·
            1 year ago

            And that is where things gets interesting. The ethics of the situation. Even beyond copyright issues. Was your AI trained on data that you have the rights for, or not?

            We then have to think of the base model. How was that trained? I have not formed a well reasoned opinion yet as to the ethics of training on social media and forum style data.

            For me, personally, I don’t have an issue with my own posts and responses ending up as AI training data. We can also argue that those posts were made on public forums, therefor in public. But does that argument hold true for everyone. Underlying that question, we have to consider the profit motif off the companies. There is a major difference between training for academic purposes and for corporate purposes.

            Valve is probably smart in steering clear of the entire mud bog at this time. Not enough is known of how it will play out in both the courts and in public opinion.

            • Ronno@kbin.social
              link
              fedilink
              arrow-up
              2
              ·
              1 year ago

              Yeah that’s what I mean, if the game devs can show that the AI language model is fully trained on its own IP, then it should be fine.