Perspective distortion reaches the model unchanged

Yes, a close, wide shot distorts proportions, and the model scores the distortion as shape. A wide lens held near a subject exaggerates whatever is nearest relative to whatever is farther away.

This is basic optics, not a flaw in any particular phone, and the model receiving the photo has no way to detect that it happened.

What focal length actually does

A wide-angle lens has a short focal length, which means it captures a wide field of view from close range. To fit that wide field into the frame, it applies more perspective exaggeration - nearer objects appear disproportionately large relative to farther ones - than a longer lens shooting the same subject from farther away would. This is the same optical effect that makes a hand held toward a wide-angle lens look oversized in a photo, or a face look subtly distorted in a close selfie taken on a phone's front camera, which tends to be wider than the rear one. Move the same camera farther back and switch to a longer focal length to keep the framing similar, and the exaggeration drops, even though the subject in front of the lens has not changed at all. The size of the effect has been modelled: Ward, Ward, Fried and Paskhover (2018), using average facial measurements in a geometric model, calculated that a photo taken 12 inches from a face makes the nose look 30% wider relative to the face in males and 29% in females, compared with a distortion-free projection, while at 5 feet there is essentially no difference.

Why the model cannot correct for it

A photograph carries no metadata about focal length by the time a scoring model sees it, and even where that metadata technically exists in the file, the model was never given a mechanism to read and apply it. More fundamentally, there is no general way to invert a perspective projection from a single image without knowing the camera's distance and focal length in advance - the information needed to undo the distortion isn't recoverable from the pixels alone, so even a model designed to try would be guessing. What reaches the scoring head is a flat image where near things look larger, and the model's job is to map that image to a number, not to reason backward about which parts of the apparent shape are lens artefact and which are the subject. This sits alongside the broader point that a photograph has no scale information at all - focal length distortion is a proportion problem layered on top of that deeper, permanent absence of scale.

How this differs from angle

Focal length and camera angle both distort what the model sees, but through different mechanisms, and it's worth keeping them apart. Angle changes which surfaces are visible and how they foreshorten - that effect is covered on its own - while focal length and distance change how strongly near is exaggerated relative to far, independent of which way the camera is tilted. A photo can suffer from one, the other, or both at once, and a model presented with the result cannot separate them any more than a person glancing at the photo could.

What this means in practice

This is a mechanism, not an instruction manual - the practical version, including how far back to stand and what lens to prefer, is a guide Rate Cock has already written, aimed at someone about to take a photo rather than someone trying to understand a result they already got. What matters here is that distance and lens choice are two of the more controllable inputs behind why the same subject can score differently across attempts, and unlike lighting, they are purely geometric - the same optical fact holds regardless of what model is scoring the image. Anyone trying to run a genuinely controlled comparison needs to hold camera distance and lens fixed across shots for exactly this reason, or the comparison is measuring lens choice as much as anything else. A human reviewer is not immune to being misled by a distorted photo either, though what a person reading a submission can factor in includes noticing that a shot looks unusually close-up, which is a compensation a fixed scoring function does not make. The practical read for a number that looks unexpectedly high or low is to check the photo for signs of a very close, wide shot before reading too much into the score itself.