That’s basically what AI is best at: incorporating lots and lots of data points (every single pixel on an image, many trillions of pixels on previous images used to train the model) to determine the best answer to a question with clear definitions of what is better or worse on a single-dimensional quality axis that most people would agree on. In this case, whether a de-noised photo looks good and accurately represents the image that could have been captured of the exact same scene using more light or a better sensor or whatever.

It’s generative AI where things fall apart. The analogy I always use is that these math techniques are great for inferring what is between pixels in a low res image, but starts going off the rails when it starts guessing what’s beyond the edges of the image. These AI algorithms are just math, and math with this many data points needs to start considering the limitations described by chaos theory (in the mathematical sense).