The original removal process requires hundreds or thousands of photos with the same watermark. Google's software can then detect the repeated image structure and then remove it completely without degrading image quality. Shutterstock's software response adds minor inconsistencies to the watermark pattern itself, using machine learning to keep it random, thus confusing Google's software. The changes are to the structure, or geometry, of the watermarks, not the opacity or location.
"The result was a watermark randomizer that our engineering team developed so that no two watermarks are the same," Shutterstock's CTO Martin Brodbek told The Next Web. "The shapes vary per image and include contributor names. By creating a completely different watermark for each image, it makes it hard to truly identify the shape." The technique is already in use, too. You can see an example of one of the new watermarks on one of Shutterstock's image pages.