Model improvement often means human eyes

Often, yes: when a policy says uploads "may be used to improve our models," a person may end up viewing and labelling your image. A scoring model does not learn its scale by itself; people assigned numbers to images upstream, and the model was trained to reproduce them.

Who is actually doing the labelling

Labelling is rarely done by the engineers who built the model. It is typically outsourced to a labelling platform or a contracted panel, working through a queue of images with instructions about what to score and how. That means an upload used for retraining is not staying inside the company that received it. It is being routed to a third party whose staff, working conditions, and data handling are set by a separate contract, and the privacy policy you agreed to is rarely the document that governs what happens at that stage.

This is a normal part of how supervised models get built, not a scandal specific to any one service. It becomes one when access is unlimited: in its May 2023 action against Ring, the FTC charged the company with "allowing any employee or contractor to access consumers' private videos" and using them "to train algorithms, among other purposes, without consent." Where scoring labels come from in the first place covers the mechanics from the model's side; this post is about what it means for the photo once a person, not a pipeline, is the next step in its journey.

What "used to improve the model" tends to skip over

A policy that says uploads may improve the product is making a true but incomplete statement. It is usually silent on whether that means automated retraining with no human in the loop, or a labelling queue where a person opens the file, or a spot-check queue for quality assurance that only some uploads enter. Those are different exposures. An automated pipeline that never shows the image to a person is a narrower disclosure than a labelling contract that puts it in front of dozens of reviewers over the life of a dataset.

Retention follows the same pattern. An image kept for training is typically kept far longer than one kept only to render a result page, because a dataset used to fine-tune future versions of a model has an ongoing reason to exist that a single scoring request does not.

What the reviewer actually sees

The image that reaches a labeller is rarely the raw file you uploaded. Most pipelines resize it, sometimes crop it to the region the task needs, and route it through a queue interface that shows one image among hundreds in a shift, with a form asking for a specific judgement rather than open-ended commentary. That framing narrows what the labeller is likely to notice or remember about any single image, but it does not remove the fact that a person, working through a screen, looked at the photo and made a decision about it.

Some labelling platforms attach instructions that forbid staff from saving, screenshotting, or discussing the content they review, enforced through contract terms and, on larger platforms, monitored access logs. Whether those terms are enforced in practice is not something an uploader can verify, and it is a different question from whether the terms exist at all. The Ring settlement is one of the few public benchmarks, requiring what the FTC called "novel safeguards on human review of videos."

What to look for before assuming the answer

A policy worth trusting on this point says, in plain language, whether uploads are used for labelling, whether that labelling is done in-house or by a contractor, and how to opt out if an option exists. The absence of any of those three sentences is itself informative: a service that never mentions human review of training data has either never done it, or has not written the policy to say so, and there is no way to tell which from outside. What an honest model card would disclose covers the more structured version of this same transparency question, for teams willing to publish one.

Consent-based commissioned review is a different arrangement entirely, closer to a service than a byproduct: what a human-reviewed judgment on Rate Penis actually involves is opt-in and the reviewer's role is explicit rather than a labelling queue three steps removed from the upload. Rate Cock states its data practices in its own policy, and that document, not this one, is the place to check whether uploads there ever reach a person outside the automated scoring path. On the measurement side, Measure My Cock's data page covers a comparable question for a tool where the underlying figure comes from a method rather than a trained model, which changes what labelling even means. Reading a tool's stated practices against its actual behaviour is its own skill, and Penis Rater's rundown of how to evaluate a scoring tool is a reasonable place to build that habit before trusting any single sentence in a policy.