Video editing applications have accumulated AI features rapidly while cameras have changed comparatively little. The gap is not about capability but about the different constraints on shipping software and shipping hardware.
Editing operates on finished material
A post-production tool works on recorded footage with no time limit, so it can spend seconds or minutes per frame.
A camera must process during capture, at frame rate, on a battery, without producing enough heat to shut itself down.
Models that run comfortably on a workstation are far outside what a handheld device can sustain, which rules out most of them at the point of capture.
Mistakes in post are reversible
An automatic edit that goes wrong is undone, because the original footage is still there.
A camera that applies processing during recording may have discarded the information needed to reverse it, and the shot cannot be taken again.
Manufacturers are therefore conservative about anything that alters what is written to the card, and confine themselves to functions that assist rather than commit.
Autofocus was the exception and shows the pattern
Subject detection and tracking are the one area where learned models have clearly reached cameras.
They fit the constraints exactly: the model runs small and fast, the output steers a mechanism rather than altering the image, and a failure produces a soft frame rather than a corrupted file.
The capability that shipped was the one whose errors were recoverable, which is the general rule for on-device features.
Product cycles differ by years
Editing software updates several times a year, and users expect the tool to change beneath them.
A camera is bought once and expected to work for many years, and firmware updates are cautious because a failed update leaves an expensive object inoperable.
A manufacturer committing to a technique in silicon is committing for the life of a product line, which is a very different decision from adding a menu item.
The division is settling into a workflow
What has emerged is capture that stays deliberately plain and processing that happens later, which is the same logic that drove the shift toward recording flat colour profiles.
Cameras compete on sensor quality, dynamic range and metadata accuracy, because those are the inputs that later processing depends on.
The practical consequence for anyone buying equipment is that capture hardware should be judged on how much information it preserves, not on what it can do to the image itself.