One possible legal theory: because the algorithm was trained on a text corpus upon which the algorithm's owner has no legal claim.

In this particular case, I don't think that theory would hold much water.

However, consider, e.g., a model that produces encyclopedia entries and is trained on a half dozen existing encyclopedias. IMO, if that model is using techniques similar to SoTA and isn't producing utter garbage, then the owner of that model should have a very difficult time claiming that the output of their model is anything more than a sophisticated round-about way of copy/pasting from existing encyclopedias.

But still, in that case, the output is still covered by copyright. It's just that the owner of the training set -- not the owner of the algorithm -- is the one with the valid claim to copyright.