So even if we’re assuming these were wholly original works that the author placed under something like a Creative Commons license, the fact that it incorporated an image they had no rights to would at the very least create a fairly tangled copyright situation that any really rigorous evaluation of the copyright status of every image in the training set would tend to argue towards rejecting as not worth the risk of litigation.
But the more likely scenario here is that they did minimal at best filtering of the training set for copyrights.