A promise with several possible mechanisms behind it

"We do not store your photos" is one sentence and several possible systems, and they are not interchangeable even though the sentence is honestly true of all of them.

The five versions

True in-memory processing. The image is loaded into RAM for the duration of inference and never written to disk at all. This is the strongest version of the claim and the least common, because it constrains the whole pipeline - logging, error handling, and queueing all have to be built to avoid ever touching disk, which is more engineering effort than most teams put in for a step that works fine either way.

Short-lived storage with a strict TTL. The file is written to disk or object storage briefly - seconds to a few minutes - to allow the processing pipeline to work, then deleted automatically by an expiry rule. This is honestly "not stored" in the sense that matters to most users, since the window during which the file exists is short enough that nothing meaningfully happens to it, but it is a different system from true in-memory handling and would behave differently under a rare failure mode like a stuck queue.

Storage of a derivative rather than the original. The photo itself is discarded, but the model's internal representation of it - an embedding, a feature vector - is kept, sometimes indefinitely, for reasons like improving future scoring or detecting duplicate uploads. The sentence is still true: the photo was not stored. Whether the derivative counts as meaningfully different from the photo is a live and separate question, since Dosovitskiy and Brox (2016) showed that "the colors and the rough contours of an image" can be reconstructed from a deep network's higher-layer activations.

Storage of the result, not the image. The score, the verdict text, and any per-axis numbers are kept as account history, which is data about the photo without being the photo. This is the version almost every tool with an account and a history page actually runs, and it is worth knowing that "we do not store your photos" says nothing about whether your results are retained, because results are a different noun.

A qualified promise that only covers the primary flow. The image is not stored by the main pipeline, but a crash report saved for an engineer to debug, a moderation queue for flagged content, or a manually reviewed edge case can still retain a copy outside the path the sentence was written about. This version is the hardest to catch from the policy alone, because the sentence is not lying - it is just narrower than it reads. Whether the brief window a file exists on disk is protected by encryption at rest, and whether the upload itself was encrypted in transit, are separate guarantees covered in encryption in transit vs at rest.

Why the sentence gets written this way at all

The phrasing is not usually chosen to mislead. Legal and product teams gravitate toward the narrowest true claim they can defend, and "we do not store your photos" is easy to defend precisely because it can be true under any of the five mechanisms above without further specification. The cost of that narrowness lands on the reader, who hears a much broader promise than the sentence was written to make, and the gap between the two is not something either side is lying about - it is a sentence doing exactly what a legal review asked of it, which is a different goal from doing what a worried user needs from it.

The one question that distinguishes them

Ask, or look for, what happens to the file between upload and the moment it is either deleted or transformed into something else. Some providers do name it: Anthropic's vision documentation, as of September 2026, says image uploads to its API "are ephemeral and not stored beyond the duration of the API request." A policy that names a duration, a mechanism, and what is retained afterwards is describing one of the first two or third versions; a policy that only makes the blanket claim is more likely running the fourth or fifth without saying so. Whether the embedding itself survives deletion of the original is the specific version of the third case worth checking on its own, since it is common and often not addressed by the "we do not store photos" sentence at all.

Reading it against a real policy

Rate Cock's privacy documentation is specific about which of these mechanisms applies rather than relying on the single blanket sentence, which is the standard worth expecting from any tool handling this kind of photo. A measurement tool faces the identical question about a photo used to derive a figure, and Measure My Cock covers its own data handling in comparable detail. A human-reviewed submission is never "not stored" in the same sense, since a person has to be able to see it to review it, which Rate Penis's judge documentation is upfront about. Comparing this specific claim across several tools' policies side by side is exactly what Penis Rater's tools coverage is for.