Nvidia recently demonstrated an impressive AI that uses machine learning to "map corrupted observations to clean signals... [sometimes] without ever observing clean ones." The AI is impressively powerful and can do many things, from reducing noise to removing watermarks.
Most examples of these methods involve training a neural network using good and bad versions of an image. What makes this technique so impressive is that in certain circumstances, it doesn't need good examples to train itself. The idea is an expansion of the concept of deviation-minimizing estimators, or M-estimators, which tell you how to estimate true data from a set of unreliable data. By expanding this idea to the training of a neural network, the researchers were able to obviate the need for clean training examples, creating an AI that can, for example, learn how to reduce noise in a photograph by looking only at noisy photographs. While there are of course specific mathematical caveats, it's a significant thing, as many applications often do not have clean examples to draw from, such as astrophotography. The scary part is that the AI can erase watermarks with ease, as you can see in the video above. Nevertheless, if the technology eventually reaches the consumer level, it could be a real boon for photographers when it comes to post-processing.