• slacktoid@lemmy.ml
    link
    fedilink
    English
    arrow-up
    15
    ·
    5 days ago

    LLM code must be GPLv3, or whatever is the most protective/restrictive license used in the training dataset code.

    • grue@lemmy.world
      link
      fedilink
      arrow-up
      9
      ·
      5 days ago

      Yep. Either that or the entire model is poisoned and nothing generated from it can be used at all.

      Either way, there is no reasonable legal interpretation where LLMs could be used to launder copyleft inputs into non-copyleft outputs.