An easier way to think about “AI” processing of the kind that’s likely being used here is that the software effectively hallucinates extra details for your image. It’s only comparing to reference data in the sense that there was a pool of many thousands photos used to train the machine learning algorithm.I wonder if this is actually simply just good quality noise reduction, sharpening, and summing of neighbouring pixels?
Neural/Machine-learning implies it is (for example) scanning the photo, working out what’s in the scene, perhaps comparing that to a reference data pool, then using that to decide the best way to process things for that particular image, even maybe redrawing elements of the image using elements from reference data, and using this process to inform future tasks.
Ultimately this is just a more sophisticated form of digital zoom. The results may look good, most of the time, if the algorithm is well designed, but there’s no way to add real details that weren’t in the original image.