The model cannot separate the subject from its presentation

Yes, grooming can move an AI score, because it changes what is visible in a photograph, not what is present. A model reads outline and contrast, grooming alters both, and a scoring model has no reliable way to tell "different presentation" apart from "different subject."

Why this counts as a confound

A confound is a variable that changes alongside the one you actually care about, without being separable from it in the data available. Hair length and density change contrast against skin, change where a silhouette's edge falls, and change how light scatters at the boundary between hair and skin - all properties a vision model reads directly, because edges, contrast and boundary texture are exactly the low-level signals earlier layers of a model respond to before anything resembling a higher-level judgement forms. The model was never given a separate channel for "grooming" versus "underlying shape." It sees one image and produces one score from whatever pattern in that image correlates with the label it learned from. Geirhos and colleagues call this failure shortcut learning (2020): "decision rules that perform well on standard benchmarks but fail to transfer" to harder, real-world conditions.

What this looks like in practice

Two photographs of the same subject, differing only in grooming, can produce two different scores on shape-adjacent axes, not because the underlying anatomy changed but because the visible boundary the model is reading changed. This is not a claim that grooming makes a score better or worse in a predictable direction - the direction depends on what the training labels happened to reward, which nobody outside the model's builders can state with confidence. It is a claim that grooming is a variable the model responds to, whether or not it was meant to be part of the judgement at all.

Why this is different from a presentation-vs-subject question

Rubrics distinguish axes about the photograph from axes about the subject as a design question, covered in more general terms elsewhere on this site. This note is narrower: even an axis intended to be purely about the subject can still pick up a presentation-level signal like grooming, because the model has no mechanism to strip it out before scoring, regardless of what the axis was designed to measure. The intent of an axis and what it actually responds to are not guaranteed to match, and grooming is one of the more common ways they diverge.

What to take from this

There is no advice implied here about grooming choices - this is a description of a scoring mechanism, not a recommendation, related to why the same photo can score differently run to run for reasons that have nothing to do with the subject changing. What follows practically is narrower: if a score changes between two otherwise similar photos, grooming is one of several variables worth ruling out before reading the change as meaningful, alongside lighting, angle and camera differences already covered elsewhere on this site. Rate Cock reporting scores by axis rather than as one total makes this easier to spot, since a shift concentrated in shape or texture axes, with others stable, points toward a presentation-level cause rather than a change in the subject itself. None of this touches actual physical measurement, which is a different property covered by Measure My Cock using a method that does not depend on a camera's read of contrast at all. A human reviewer notices grooming consciously and can say so in a sentence, which is a structural difference from a model folding it silently into a number - how a human judge's process differs from a model's is worth reading alongside this. Ruling out variables like this before trusting a score change is the same habit Penis Rater recommends for reading any single result.