What you can test without seeing the training set

You do not need access to a model's weights or training data to find out whether it is biased. You need pairs of inputs that differ in exactly one attribute, and a way to compare what comes back. This is counterfactual auditing, and it works on any scorer you can only reach through an upload box.

The core idea

Pick an attribute you suspect matters to the score: skin tone, grooming, background clutter, camera distance. Build or find two images that are as close to identical as possible except for that one attribute. Submit both, repeatedly, and compare the distributions of scores rather than single results. If the two distributions differ by more than the tool's own run-to-run noise, the attribute is doing something the score was not supposed to be about. If they overlap heavily, that particular attribute is probably not a live source of bias, for this tool, on this axis.

The logic is the same one behind a controlled experiment in any field: hold everything constant except the variable under test, and attribute the difference in outcome to that variable. What makes it an audit rather than a general controlled comparison is the specific goal - you are not asking whether the tool is consistent, you are asking whether one particular attribute moves the number when it should not.

Building the pairs

The hard part is not the statistics. It is constructing pairs that actually isolate the attribute you care about. A real photo pair rarely varies in only one dimension - change the lighting to test grooming and you have also changed the lighting. Some auditors use image editing to alter a single attribute directly on the same base photo, which controls for everything else at the cost of introducing editing artefacts the model might react to instead. Others use large sets of naturally varying photos and control statistically, adjusting for the attributes they can measure and hoping the ones they cannot are evenly spread across both groups. Neither approach is clean. Editing risks the model reacting to the edit rather than the attribute; natural pairs risk confounds nobody thought to check for. A serious audit usually runs both and looks for the same direction of effect in each, since agreement between two flawed methods is more convincing than either alone.

What counts as a finding

A single pair proves nothing - it could be the ordinary variance any repeated score carries. The spread across repeats has to be measured first, on the same tool, so you know what a difference has to clear before it means anything. Tracking your own scores across submissions runs into the same noise-floor problem from the user side, which is the practical version of this that penisrater.com walks through. A finding is a consistent gap across many pairs, in one direction, larger than the noise floor. This is also why single anecdotal reports of bias ("my photo scored low and I think it's because of X") are weak evidence on their own - they are single-pair audits with no noise estimate, and the tool's ordinary variance can produce a low score on any given run for no systematic reason at all.

What the method cannot tell you

Counterfactual auditing finds correlation between an attribute and the score, from outside the model. It cannot tell you the mechanism - whether the model learned the bias from labeller behaviour, from an imbalanced training set, or from something more subtle in how images were preprocessed. Where the labels came from and what the training set over- or under-represented are separate questions an external audit cannot answer directly, only motivate. It also cannot rule out bias in attributes you did not think to test. An audit is only as good as the list of attributes someone bothered to check, and a clean result on the attributes you tested says nothing about the ones you did not.

Where this method is actually used

Fairness researchers use versions of this technique on closed commercial models routinely, because it is often the only tool available when a company will not publish its training data or evaluation breakdowns. The same logic applies to a rating tool: pairing across skin tone is the standard way researchers test exposure and skin-tone bias in vision scorers, and it generalises to any attribute a rubric was not supposed to be scoring. Rate Cock publishes its per-axis breakdown on every result, which is what makes this kind of external audit tractable in the first place - a single blended total gives an auditor nowhere to look, while separate axes let a systematic shift on one axis show up against the others staying flat. A human panel does not remove the need for this kind of check either; what a panel of raters brings and where its own biases sit is a related but distinct question, since a group of people can share a bias as easily as a model can learn one. None of this requires size data or measurement claims to run - the whole method operates on the score alone, which is also why measuremycock.com's data-side work sits upstream of it rather than overlapping with it.