The training data would have contained millions of copyrighted images. So yes absolutely. And what its doing there with the watermark looking thing is outputting a ballpark representation of many original images from its training data that just so happened to contain a watermark, probably shutterstock or something similar.