Models pick up age-correlated cues and score them

Nobody sets out to build a scoring model that responds to age. It happens anyway, because skin texture, tone and elasticity all correlate with age across a broad population, and a model trained to notice texture and tone has no way to avoid also picking up the correlation that rides along with them.

How a proxy gets learned without being labelled

A model trained on images and scores never sees an "age" label unless a dataset specifically included one. It learns whatever pattern predicts the score it was shown, using whatever is visible in the pixels. If the people who produced the training labels were, even slightly and unconsciously, responding to age-correlated cues when they scored images - and there is no reason to assume they were not, since age-correlated skin properties are visually salient to anyone looking at a photo - then those cues become predictive of the score during training, and the model learns to use them, whether or not "age" was ever named as a variable to anyone involved. This is proxy learning: the model was never told to look for age, but skin texture and tone were available, correlated with the label, and therefore useful, so it uses them.

Why this is hard to separate from the model's other behaviour

A model that has learned an age-correlated pattern does not announce it. The output is still a single number on a rubric axis, indistinguishable in form from a score that responds to something else entirely. There is no flag reading "this score was partly a function of an age-correlated skin cue" anywhere in the interface, and there is no reliable way for a user to detect it from a single result, because the pattern is baked into the same weights that produce every other part of the score.

This connects to, but is distinct from, the exposure and dataset-composition mechanism covered for skin tone specifically - that piece is about camera calibration history and dataset representation across skin tones. This one is about a different mechanism: a correlation the model absorbed between an age-adjacent visual property and the score it was trained to predict, which can exist even in a dataset with even representation across skin tones, because the correlation runs through texture and tone as they vary with age within any given tone, not between tones.

Why this is not a claim about accuracy in either direction

It would be a fabrication to say this makes scores systematically higher or lower for any age group - that depends entirely on what the underlying human labels happened to reward, which is not published and not knowable from outside. What is defensible is the mechanism, not a magnitude or direction: age-correlated visual properties exist in skin, they are the kind of property a texture-sensitive vision model reads easily, and there is no design step in a typical training pipeline that isolates and removes this correlation unless someone specifically built one in. Absent evidence of that specific mitigation, the default assumption for any texture-sensitive scorer should be that some amount of age-proxy learning is plausible, not that it is absent.

Why this differs from an intentional age-estimation model

Purpose-built age-estimation models are trained explicitly on age-labelled data to predict age as their stated task, and their accuracy and failure modes are a distinct, separately studied question. What is discussed here is the opposite situation: a model with no age-estimation task at all, trained only to predict a score on an unrelated rubric, still ending up encoding age-correlated information as a side effect of what predicted the training labels well. The absence of an explicit age task does not mean the absence of age-correlated behaviour - it just means nobody deliberately built it in, and nobody deliberately tested for or removed it either, in most public scoring tools.

What this means for reading a score

Not much changes practically, because there is no visible signal to act on - this is a description of a plausible mechanism inside the model, not a symptom you can spot in a single result, the same limitation covered in what a model is actually doing when it produces a number. What it argues for is the same scepticism this site recommends generally: treating a score as a read of the photograph's low-level properties, filtered through everything the model happened to learn from its training labels, rather than as an assessment of the subject in any clean sense. Rate Cock reporting scores per axis at least narrows where an age-correlated effect, if present, would concentrate - texture and surface-condition axes are the more plausible carriers of this kind of proxy than a size-adjacent axis would be. None of this applies to physical measurement, which Measure My Cock treats as a data question separate from any model's learned correlations. A human reviewer can also carry age-related assumptions into a judgement, through an entirely different, more conscious mechanism, and how human judges are trained and calibrated is worth reading as a genuinely different failure mode rather than an automatic fix. Reading a rubric axis for what it is actually likely to be responding to, rather than what its name implies, is a habit Penis Rater covers in more general terms.