Metadata is written by the person who needs it least

Read the standard advice and it arrives as a policy question. Decide who tags. Decide when. Write it down. Enforce it at upload.

Every version of that advice assumes the same thing: that the right tags exist and someone is failing to write them down.

That is a reasonable answer to a different problem. It treats tagging as a discipline failure, as though the work is simply not being done. The work is being done. It is being done by someone with no way of knowing what will be asked.

Knowledge and need sit in different places

At upload, one person holds the full context. The shoot, the brief, the reason a frame was chosen. They tag what seems relevant, which means what is salient to them, that week, for that campaign.

Nine months later somebody else needs an image. They are not looking for the campaign. They are looking for something with an autumn palette, or a person over sixty, or a location that reads as northern Europe. None of those were relevant at upload, so none of them were written down.

The failure is not laziness. It is that relevance at capture and relevance at retrieval are two different things. Only one of them is knowable when the tagging happens.

What a policy actually produces

Policies close the gap by asking for more. More fields, filled more consistently, checked more often.

What that produces is tagging that satisfies the policy. A required field gets a value because the upload will not complete without one. The value is true and generic, because generic is what you write when you do not know what the question will be.

Then volume arrives. A shoot produces two thousand frames. A policy that works for twenty does not survive two thousand, and what happens next is predictable. The fields get filled once, in bulk, with whatever is true of the whole set. Structure that describes everything describes nothing in particular.

Metadata is a prediction, not a description

This is the part the policy framing hides. Tagging is usually described as recording what an asset is. It is not. It is a bet on what someone will one day ask, placed by a person who cannot see the question.

Read that way, the conclusions invert. Asking people to try harder does not help, because effort was never the constraint. Foresight was, and nobody has it.

What survives is the structure nobody had to predict. Facts the system can read because they are already in the file: when it was taken, with which camera and lens, in what form, at what size. And description generated from the image rather than remembered about it, which has the useful property of not depending on what the uploader thought mattered.

That does not make human tagging pointless. It moves it to where it is genuinely irreplaceable, which is intent. A machine can say there is a person on a bicycle. Only someone who was there can say the shot was chosen because the light held for four minutes.

Where this applies

A library one person fills and one person searches does not have this problem, because both people are the same person. Neither does a small set where scanning beats searching. It matters at volume, across time, and across teams. By then the uploader has moved on, the campaign is over, and the asset is being considered for something nobody imagined when it was filed.