In the example above, the result looks pretty clear, but I tested it on some other images, and the results were not so definitive. It seems like interpreting the output is still going to be a challenge in some cases. Also, I am afraid that maybe it can be defeated by just taking a screen shot of the edited image and then inserting the screenshot?

You should avoid using it with screenshots scanned images of documents or juxtaposed images that are in fact altered images.

Still, any bit of technology that adds friction to the abuse process is an improvement. Thanks for the tip!