Compression artefacts are features, as far as the model knows

Yes, heavy JPEG compression can change an AI score, because the blocking and ringing it leaves behind are texture to a vision model. A model reads texture directly off the pixels, and compression artefacts are indistinguishable in kind from the surface detail it was trained to score.

A cheap phone camera under poor light can compound this, since sensor noise gets compressed alongside real detail, and how camera quality itself biases a score is a related but separate mechanism. Researchers have tested this directly: Dodge and Karam's 2016 study of image quality and deep networks tested JPEG and JPEG2000 compression alongside blur, noise and contrast, and found networks "susceptible to these quality distortions, particularly to blur and noise," with compression among the effects tested.

What compression actually discards

JPEG works by dividing an image into small blocks, typically 8 by 8 pixels, and representing each block's content as a sum of frequency components, then discarding the higher-frequency components more aggressively as the compression level increases. High frequencies are where fine detail and sharp edges live, so this is where quality loss shows up first. At low compression, the loss is close to invisible. At higher compression, two artefacts become visible: blocking, where the edges of the 8x8 grid become faintly visible as the blocks stop blending smoothly into each other, and ringing, faint echoing lines that appear near sharp edges where the frequency approximation overshoots. Both are systematic and repeat in a recognisable pattern across the image, which is exactly the kind of regular, learnable pattern a vision model's early layers are good at picking up, in the same way they pick up on edges and low-level structure generally.

Why this matters for a texture-sensitive axis

A model scoring surface condition or overall visual quality is, at some level, reading local texture statistics - contrast, sharpness, the presence or absence of fine detail. Compression artefacts change all three, and they change them in a way that has nothing to do with the subject. A heavily compressed photo of good, clear skin can present degraded local texture that looks, numerically, closer to what the model associates with poor image quality, simply because the compression damaged the fine detail the axis relies on. This is a distinct mechanism from resolution loss - downscaling a photo discards detail by shrinking the image before the model ever sees it, while compression discards detail within the pixels that remain, at whatever size the image already is.

Recompression through messaging apps

A photo rarely reaches a scoring tool as the single, original file. Model providers flag the same risk; Anthropic's vision documentation warns that lossy compression "can introduce artifacts that are detrimental to model performance, especially when multiple compression passes are applied." Sending a photo through a messaging app typically recompresses it to reduce bandwidth, and if that file is then saved, forwarded, and sent again, each pass can compound the artefacts, since JPEG compression is not fully reversible and re-encoding an already-lossy file loses more each time. A file that has been through several rounds of app-to-app forwarding can carry noticeably more artefacting than the original camera output, even if nobody involved intended to degrade it, and even though the visible image looks roughly the same to a person glancing at it on a phone screen. Saving a photo out of a messaging app and re-sharing it elsewhere repeats the whole cycle again, so a file's history of hops is a rough proxy for how much of its original detail actually survives.

What this means in practice

This is a mechanism worth knowing, not a size question - it has nothing to do with what a photo can or cannot tell a model about measurement, which is a separate and permanent limitation unrelated to file format. The practical upshot is that the file you actually upload matters: a fresh export from the camera, sent as a file rather than passed through several rounds of chat compression, gives the model cleaner texture to work with, and that is part of why the same subject can score differently across attempts if the file source varies between them. A tool that lets you re-upload the original rather than a screenshot or a much-forwarded copy is giving the model a fairer shot at reading the actual surface, a distinction related to but separate from how a screenshot changes the file in its own ways. Rate Cock scores whatever file it receives, which is one more reason to send the original rather than a copy that has already been through several apps, and it's the kind of detail worth checking before trusting a comparison between two results. A human reviewer is comparatively unbothered by mild compression - a person reading a photo tends to look past minor artefacting the way they look past a slightly grainy phone photo, which a model has no equivalent instinct to do.