Not a lawyer but I imagine the infringement on the input side, during the dataset creation and training would take place wherever the model is being trained? So presumably the us? On the other hand, yeah. Memorizing input data and spitting out protected characters on the output side seems like an issue.