> AI-assisted autofocus, subject tracking...

This is already present for the last decade. It's the foundation which makes "sports DSLRs" possible. AI here is not what we know as today's "AI". These are specialized models baked into DSPs to run at inhuman speeds. In Sony's case it's phase data, image itself, information from the lens, and sometimes the accelerometers on camera. This is not making something up from thin air.

AI assisted autofocus is smarter phase detect autofocus. Subject tracking is an educated guess for where the focus shall move, based on movements in the image.

My ~20 year old Nikon D70s has a 1004 pixel "scene detection and metering" sensor "trained" on thousands of images to make educated guesses in full-frame metering mode to exposure correctly. This is in no way helping or altering the image in any way. Same for faster autofocus and object tracking.

Exposure, white balance, noise reduction, color correction, HDR et al can be all achieved with deterministic and traditional algorithms as well. Some of these processes have history in film photography and carried over to digital from analog. Some photographers are famous for their darkroom treatment of their shots during development.

Processing an image with deterministic algorithms using the data in the raw photo is fine. This is how darkrooms work as well. However, guessing things, altering images by adding something which was not there and claiming otherwise is not acceptable. This can be a new element inside the scene or pixel details which was not resolvable by the lens/sensor/film to begin with. Both are "photomanipulation", not photography.

I feel like the distinction comes down to "edited" vs "generated", but even here the line can be blurry. For example, is AI denoise really just removing noise from an image, or is it generating a new image of what it thinks the original signal was?