Accuracy
A screenshot is a different image
A screenshot re-samples the display, changes resolution and colour, and drops metadata; the model sees a different file and may score it differently.
Guides on Accuracy: There is no true score to be accurate against, Change one thing, hold the rest, repeat, Exposure that suits one skin tone hides another
Yes, it matters: a screenshot of a photo is not the photo. It is a new image, captured from your screen by the device's own capture tool, and the model scoring it sees that new image, with its own resolution, colour handling and compression, not the original file.
What actually changes
Taking a screenshot re-samples the original at whatever resolution the display and the crop happened to produce, which is rarely the same pixel dimensions as the source file. That resampling is a resizing operation in its own right, distinct from and on top of whatever downscaling the model's own preprocessing does later. Colour can shift too: the screen renders the file through its own colour profile, and the screenshot captures the rendered result rather than the file's original colour data, so a screenshot can carry a subtly different tone than the source ever had. The screenshot is then usually saved as a fresh PNG or JPEG, with its own compression applied on top of whatever compression the original file already had - a second, unrelated pass of exactly the kind of artefact-introducing process covered in more detail elsewhere.
Zoom level matters too. If the photo was viewed zoomed in or cropped tightly on screen before the screenshot was taken, the model receives a different framing than the original photo had, which changes how much of the frame the subject occupies - a variable covered on its own that a screenshot can introduce by accident.
Metadata, briefly
A screenshot also does not carry the original file's metadata - it is a new file, written by the operating system's capture tool, with its own fresh metadata rather than the source photo's. That side of the question belongs to the privacy read of what a screenshot does and does not remove, and it is a separate concern from anything about how the model scores the resulting pixels.
The practical read
None of this means a screenshot is unusable, only that it is a measurably different input, and a model has no way to know or correct for the fact that it is looking at a re-sampled, re-compressed rendering rather than a source file. Image classifiers are known to be sensitive to exactly this kind of degradation: Dodge and Karam (2016) tested four networks against blur, noise, contrast and JPEG compression and found them "susceptible to these quality distortions, particularly to blur and noise." If you have the original file, uploading it rather than a screenshot removes one more variable from the spread you'd otherwise see across repeated attempts. This has nothing to do with what a photograph can tell a model about physical size either way - that limit is present in both the original and the screenshot equally. Rate Cock scores whatever file arrives, screenshot or original, and the difference is invisible from the result alone, which is exactly why it's worth knowing about before comparing two scores against each other. A commissioned human reviewer is far less sensitive to this kind of file-level difference than a model is, since a person is reading the image the way it looks on their screen either way, largely unbothered by the resampling underneath it.