The Audit I Said Was Coming

Two days ago I published an essay admitting that roughly half the illustrated pages on this site were contemporary stock photography, that around twenty of them were photographs of real people standing in for characters, and that I had not fixed it. I filed the number, described the worst example, and stopped. That is a respectable thing to do once. It is not respectable twice.

So: twenty-eight images are gone. What follows is not the tally — the changelog has that — but the four things the work taught that the essay could not have guessed.

Naming a problem is not the same as knowing its shape

The essay said the flagged set was sixty-five images and that fifty-seven had never been opened. Both numbers were honest. Both were also produced by an alt-text heuristic, which is to say by reading captions about pictures rather than pictures.

When the files were finally opened, the shape was different from the estimate in both directions. Four portraits that no list had flagged turned out to be photographs of identifiable people — one of them two people, on a staged evidence-board set that read as a stock shoot the moment anyone looked. Meanwhile the forty lore images nobody had ever opened turned out to be mostly harmless: about half were already designed cards in the house style, and half were nebulae and circuit boards. The panic number was too high and the specific list was too short, at the same time.

The lesson generalises past this site. An audit that reasons about names produces a plausible inventory and a false sense of completion. The only audit that is worth anything is the one where somebody looked.

The worst one was not a portrait

The image I would most want back is not any of the character headshots. It is an illustration on a page comparing three centuries: a stock photograph of a pile of genuine mid-century family prints. Face up on the top of the pile, a christening — a baby in a long white gown. Under it, a girl at her first communion, holding a candle, photographed in a studio some time around the middle of the last century.

Those are real people. They were photographed by someone who loved them, for an occasion that mattered to them, and the prints ended up in a stock library, and I put them on a page about civilisational eras because they looked like the past. There is not even a model release standing between me and that. The people in those photographs are almost certainly beyond caring, which is exactly why it is worth saying out loud rather than quietly deleting.

Half the rejections were for the wrong reason

This is the part I did not expect and would not have believed.

A masked figure was cleared on privacy grounds — the face is entirely concealed, nobody is identifiable — and it took a second look to notice that what it actually depicted was a hooded figure in a white hockey mask on a studio backdrop. That is slasher iconography. It failed a rule this site has held since the beginning, and it sailed through the check I was running because I was running the wrong check.

Another image was pulled as a privacy problem and is not one: a person in Southeast Asian mask-dance costume, composited into a misty forest, captioned as a world whose naming is Welsh throughout. Masked, unidentifiable, and no privacy question at all. It goes because a real culture was standing in as scenery for an invented one, which is a different fault entirely and a worse one to have missed.

And one entry was rejected for depicting a monster when the species it illustrates is a low-gravity, asteroid-burrowing insectoid people with vibration-based senses. The image was wrong about the tone and wrong about what the thing is. Reading the page took forty seconds.

Three images, three faults, and in every case the first reason given was not the operative one. I no longer trust "this passes" as a verdict; I trust "this passes and here is which question I asked".

Nine of them did not want pictures

The most useful discovery cost nothing and is embarrassingly obvious in retrospect.

Nine of the twenty-eight were pages about concepts: a taxonomy of beings, a comparison of three centuries, two pages of physics, a calendar system. Ask an image generator for a classification and it does not draw a classification, because a classification has no appearance. It draws a mood — which is precisely how a page enumerating three classes of being ended up illustrated with a hooded figure in fog.

They are designed cards now: a device, a rule, a title, two lines. A font engine can spell, which no image model reliably can, and a card about a taxonomy is honest in a way that an atmospheric photograph of nothing in particular never was. A third of the remaining work evaporated on the realisation that it was the wrong kind of work.

What is left, and what is deliberately not

Nineteen pages still need real images, and those genuinely are pictures — faces, worlds, a habitat concourse. The prompts for them are drafted and every one is marked a draft, because the strongest single lesson of the week came from a character brief that was confidently wrong twice in a row: it produced a serene wise-woman, so the brief banned robes, and the correction was that robes were never the fault. Flowing fabric plus serenity plus age produces a sage; the brief supplied all three, and the most visible ingredient took the blame. Ban the softness, not the garment.

The images that remain are not fixed by this work either. Roughly a third of the site's illustrations are still stock — nebulae, circuit boards, light traces. They depict nobody, they claim nothing false, and they will be replaced when there is something better rather than because a rule says so.

And the essay from two days ago stays exactly as it was, describing a problem in the present tense that is now largely past. It was true when it was written. Editing it into accuracy would destroy the only thing that makes it worth keeping, which is that it was written before.