while it's possible that the courts could come up with a 3rd option, the vast majority of the analysis I've seen results in this going one of two ways:

• LLMs can copywash a license away, in which case there's nothing from preventing a "clean room" LLMwashing of things like MS Windows source-code or Photoshop; and at least in the US, the output of LLMs has been determined to not be copyrightable, so effectively public-domaining any existing source-code it (re)"creates", infuriating big companies that rely on copyright (software, publishing, music, etc)

• LLMs can't copywash, so original licensing still holds. LLM output that effectively reproduces copyrighted material would likely result in lots of verdicts of infringement against modern AI companies