Enlarging a low-resolution image with a modern upscaler produces convincing detail: individual hairs, fabric weave, readable texture. None of that detail came from the original file, and the distinction has practical consequences.
The information is genuinely absent
A small image contains a fixed number of measurements of the scene, and detail finer than that spacing was never recorded.
Conventional enlargement interpolates between the values it has, which is why it produces a soft image rather than a detailed one.
Softness is an honest representation of the situation: the detail is not there, and interpolation does not claim otherwise.
Learned upscaling generates rather than interpolates
A model trained on pairs of high and low resolution images learns what fine detail typically accompanies a given coarse pattern.
Given a new low-resolution input, it produces detail consistent with that learned association, which is a generated guess rather than a recovery.
The guess is usually right in character and never guaranteed right in specifics, since many different scenes reduce to the same coarse image.
The failure mode is confident and specific
Upscaling a blurred face produces a sharp face, and it is a face the model considered likely rather than the face that was photographed.
Unreadable text becomes readable text, with letters chosen for plausibility, which can produce a licence plate or serial number that looks definitive and is fabricated.
This is why forensic practice treats upscaled imagery as unusable and works from the original, however poor.
Quality metrics reward the wrong thing
Perceptual scores measure whether an image looks like a natural photograph, and generated detail scores well on exactly that basis.
Measures of fidelity to the true scene are lower for such output than for plain interpolation, because invented detail is wrong in a way that blur is not.
Products are tuned for the perceptual measure, since that is what users respond to, and the trade is rarely stated in the interface.
Where the technique is entirely appropriate
For presentation work — enlarging an image for a print, restoring an old photograph, preparing footage for a larger display — plausible detail is the goal and there is nothing to misrepresent.
Restoration of damaged archival material sits at the boundary, and archives handle it by retaining the original alongside the enhanced version with the processing recorded.
That convention is the useful one generally: the enhanced image is a rendering, the original is the record, and problems arise only when the rendering is filed as the record.