Why age of the image matters

An old photo sitting in a camera roll feels like the same kind of thing as a photo taken this morning. For scoring purposes it is not. Every variable that a comparable series depends on holding still has had years to change, and an old file changed none of them on purpose.

What has drifted

The device is probably different, and a different phone means a different lens and different processing, covered in more depth in why the same device matters across a series. The conditions - light, distance, angle, background - were whatever they happened to be that day, not a protocol anyone was following, because nobody photographs casually with future comparability in mind. And the subject's own state at the time is simply a different data point than its state today, not a flaw in the photo but a fact about what it can be used for.

None of this makes the old photo bad. It makes it ineligible for the specific comparison people reach for it to make - "here is what I scored then, here is what I score now" - because that comparison needs everything except the thing being tracked to have stayed constant, and nothing did. Model accuracy itself also shifts over time as tools get retuned, which is one more reason an old score and a new one are not measuring the same thing even before the photo's own drift is counted. That shift is measurable: Chen, Zaharia and Zou (2023) found GPT-4 identified prime numbers with 84% accuracy in its March 2023 version and 51% in its June 2023 version - the "same" service, three months apart.

What it is still good for

A standalone data point, read on its own terms, without weight put on the gap between it and a current score. It is also useful as a test of the tool itself rather than of anything else - resubmitting an old file to see how a tool's own calibration has shifted is a legitimate use of exactly this kind of image, because there the age is the point rather than a confound.

The honest fix

Start a new baseline today, under recorded conditions, and let comparisons run forward from there. An old file cannot retroactively become part of a controlled series; it can only be filed as what it is, a single unlogged sample from a different time.

That is a smaller inconvenience than treating a several-year gap in device, light and state as noise and drawing a conclusion from it anyway, and it applies whether the second opinion comes from a model or from a person - either way, an unlogged old photo and a logged current one are not the same kind of evidence. It is also the same discipline measurement practice insists on: a number without its conditions attached is not comparable to anything, no matter how the number was produced, whether that number sits on Rate Cock or anywhere else.