Hands were the signature failure of generated images and are now largely handled. Reflections were a quieter failure and remain broken. The difference explains what these systems can learn.
Hands were a data and emphasis problem
Hands are complex, articulated and appear at many scales, but they are a bounded object with consistent structure. There are five fingers and the joints bend in known directions.
Training material contains an enormous number of hands, and once the failure became notorious, data curation and evaluation began targeting it directly.
A problem that responds to more examples and more attention on a well-defined object is the kind these systems solve, and the improvement arrived accordingly.
A reflection is a function of the whole scene
What a mirror shows depends on the geometry of everything in front of it and on where the camera stands. It is not a property of the mirror.
Rendering it correctly requires a consistent three-dimensional model of the scene, including objects outside the frame that would nonetheless appear in the reflection.
Image systems do not build that model. They produce a plausible arrangement of pixels, and plausible is not sufficient when correctness is determined by an off-screen relationship.
The failure is systematic rather than random
Generated mirrors typically show a nearly correct scene with the wrong details: a subject facing the wrong way, an object present in one and absent in the other, a room that does not correspond.
This is what a system optimizing for local plausibility produces. Each region looks like a reflection should look, without any constraint tying it to the source.
Storefront windows, sunglasses, still water and polished floors produce the same class of error for the same reason.
Evaluation did not push on it
Hands got fixed partly because their failures were obvious to everyone, including people with no interest in the technology. That made them a visible target.
Reflection errors require a viewer to compare two regions of an image deliberately, and most viewers do not, so the failure generated far less pressure.
Where nobody is measuring a property, the systems do not improve on it, which is a general pattern rather than a fact about mirrors.
Practical work routes around it
Retouchers handling generated images for American commercial clients treat reflective surfaces as manual work rather than as something to prompt for.
Common approaches are to compose the shot without a reflective surface in frame, or to generate the reflection separately and composite it deliberately.
The tell survives as a reliable indicator, which is why people examining a suspect image check the glass before they check the hands.