Are "open" models exempt from that? Can imagine they used the same datasets.

If they provably shared those datasets then they're just as liable for piracy.

Many of the open weight models are trained on outputs from these models (distillation)

Most open models are developed for profit so they should be equally liable.