What the GenAI Machine Has Never Seen

Generative AI models or "GenAI" learn from the average. They train on billions of image-caption pairs scraped from the internet, and what they extract isn't the exceptional but the statistically frequent. Which makes them extraordinary at producing what they've seen most — the cinematic poster, the corporate portrait, the cover landscape — and notably poor at the rare, the specific and the genuinely idiosyncratic.

That isn't a temporary limitation waiting on more parameters. It's a direct consequence of how they work.

And it opens a commercial window worth looking at carefully.

The data behind GenAI

One study compared the stereotype scores these models produce for different nationalities against each country's number of internet users. The result was consistent: the smaller a place's digital footprint, the more stereotyped the image the model hands back.

Translated: the under-photographed gets represented badly. Not out of malice — out of missing material.

That finding has an optimistic reading almost nobody is making. If the structural weakness of these tools is everything for which few images exist, then visual scarcity has just become a competitive advantage. What is poorly documented is exactly what can't be generated.

The cost of looking like everything

The flip side of the same phenomenon is homogenisation.

Thousands of brands and creators using identical tools produce interchangeable material. There is now a recognisable visual category people simply call "AI," and landing inside it carries a measurable cost. A viewer who identifies an image as generated reacts fast and ungenerously: assigns it less effort, less credibility, and scrolls on. The insult attached to the low end of that output — slop — describes the reaction well. It short-circuits the aesthetic experience instead of rewarding it.

Algorithms make the loop worse. They don't reward the original, they reward what already generates engagement, which tends to be the familiar. So with more creation tools available than at any point in history, visual diversity is narrowing rather than widening.

The obvious play

If the system is strong where material is abundant and weak where it's scarce, the move is toward scarce.

In practice that means concrete, unglamorous things. Photographing places that don't appear in a search engine's first hundred results. Working with people who don't fit the casting archetype. Documenting trades, geographies and communities with little serious visual record. Returning to one subject over years instead of touching it once.

None of that is a moral position. It's where the value sits once everything else gets cheap.

What it does to the archive

If this holds, it changes the calculation of what's worth keeping.

For twenty years archive logic was subtractive: too many images, not enough storage, delete whatever didn't get published. That logic assumed the value sat in the individual photograph and its immediate use.

Under the opposite argument, the value sits in density. A hundred images of an under-documented place, made across years, with dates, locations and context recorded, are worth considerably more than a hundred good images of a hundred different places. Not because any one is better, but because together they constitute something that doesn't exist anywhere else.

The consequences are deeply unromantic. Metadata matters. File naming matters. Recording who appears, where and when matters, and it's boring. A disorganised archive of irreplaceable material is, practically speaking, half an archive.

It also changes what gets signed. Documented provenance and consent in order stopped being paperwork and became part of the asset, because verifiable material is scarce and getting scarcer.

The counterweight

None of this solves the real problem, which is still distribution.

Being rare is worth nothing if nobody arrives. Algorithms reward the familiar, and the seldom-seen is by definition unfamiliar — so the same quality that makes a body of work valuable makes it hard to circulate. That's the price of working away from the centre, and it hasn't gone anywhere.

What's different from three years ago is where the bottleneck sits. It used to be competing against other photographers for the same commission, in the same place, for the same possible image. Now generic competition is infinite and free, which paradoxically makes it easier to be distinctive and harder to be seen.

Trading a production problem for a distribution problem isn't a victory. It isn't a defeat either. It's a different problem — and this one can actually be worked.

The other scarcity

The argument so far has centred on places, but the scarcity hardest to replicate isn't geographic. It's access.

A generative model can produce any room you like. It cannot get someone to let it into theirs. It cannot sustain a ten-year relationship with a community, or earn permission to be present at a moment nobody else will see, or build the trust that makes a person drop their guard in front of a camera.

That last one is probably the most underrated skill in the profession right now. There's endless conversation about technique and almost none about the capacity to be in the right place with permission, which is a social skill and takes years.

And it has an interesting property: technology doesn't make it cheaper. The reverse, if anything. The easier it becomes to fabricate an image of someone, the more an image made with someone is worth.

The caveat

Worth saying plainly: rarity on its own isn't a virtue. Photographing something under-photographed and doing it badly is still doing it badly.

But the combination of real access and real craft has become hard to replace in a way it wasn't three years ago. For a long time, working far from the centres of image production was a distribution handicap. Now it's a data advantage.

An unromantic way to put it. Also the first piece of structural good news the profession has had in a while.

Sources: OASIS, analysis of stereotypes in text-to-image models · Kompozy, The AI Design Aesthetic · Medium, The Slop Aesthetic