This. It's also while you'll sometimes get a mangled Getty Images watermark on some image generations, or a logo in the bottom left corner. If it's a prominent feature in the training dataset it'll show up, exactly how these models are supposed to work.

The 'bug' here is whatever post-processing step or system prompt is in place to steer the model away from doing this.

Maybe it would have worked better to preprocess the training data…

Maybe don't use stolen images in the first place.