What AI
Learned to See

When an AI generates an image that looks like a photograph, I find myself asking a simple question. Where did it learn that?

The answer is not complicated. It learned from photographs. Hundreds of millions of them. Images made by cameras, by photographers, over nearly two centuries. It looked at that body of work and found patterns. The quality of light through glass. The compression of distance that a long lens produces. The texture of film grain. The tonal relationships that a sensor or an emulsion records. It absorbed these not as choices, but as facts: because in photography, they are facts. They are the direct result of how the medium works.

This is the inheritance. Before AI could generate an image, photography had to exist to define what an image looks like. The box, the lens, the light, the chemical reaction, the chip. These are not aesthetic preferences. They are the technical architecture of seeing. And that architecture is what AI was trained on.

I think about this often in the context of Zero Baseline. The platform maps photography’s origins through first photographs. Every entry traces a moment when the medium discovered something it could not yet do before. A new sensitivity to light. A new way of capturing motion, or depth, or colour. The technical genealogy follows a clear arc: chemical process, then digital sensor and chip, then file format, then computational photography, and data visualisation. Each step moved further from the direct relationship between light and surface. But each step, until recently, remained grounded in physical measurement. That grounding is what AI breaks away from.

When the Root Is Lost

In 2024, researchers at Oxford and Cambridge published a paper in Nature that analysed what is known as model collapse. When AI models are trained on data that is itself AI-generated, rather than on original human-made content, the models degrade. Progressively and irreversibly. The range narrows. The edges of the distribution disappear. Within a few generations of recursive training, what began as rich and varied output becomes a statistical average with no memory of what it was averaging.¹

The paper, authored by Shumailov et al., is precise: indiscriminate use of model-generated content in training causes irreversible defects in the resulting models. What goes first is everything at the edges: the rare, the specific, the particular.²

Applied to images, the implication is significant. If AI continues to train on AI-generated images, the inherited photographic visual language, all that grain, light, colour, and depth accumulated across a century and a half of technical evolution, begins to disappear. Not replaced by something new. Replaced by a flattened average with no connection to its origin.

Two Kinds of Seeing

It is worth pausing here on what data visualisation actually means. Not all computational images are equal.

When the Event Horizon Telescope Collaboration published the first image of a black hole in 2019, what the world saw was data visualisation. Radio telescope measurements gathered from observatories across the planet, processed and rendered into a visible image. The process was computational. But the data came from physical reality: from actual measurements of an actual phenomenon at a specific location and time. Filmmaker and Harvard historian of science Peter Galison documented what happened next in his 2021 film The Edge of All We Know. The collaboration divided into independent imaging teams, each given the same data set and kept separate from the others. Each team worked through the same scientific process without knowledge of what the others were producing. When the results were compared, every team had arrived at consistent images. The image is not arbitrary. It is created by the data.

NASA operates on the same principle. Its image archive is among the most openly documented in the world. Original file data, instrument metadata, mission nomenclature: all accessible, all traceable. Behind each image is a documented process involving many people, many instruments, and measured physical data. When NASA visualises data, it is showing us something that exists.

AI image generation is technically also a form of data processing. But the data is statistical patterns drawn from other images, not physical measurements of the world. There is no instrument. There are no coordinates. There is no moment of capture. The result may look like a photograph. It is not a record of anything that happened.

This is the distinction that matters. And it is one that will become harder to maintain as computational and AI-generated images increasingly share the same platforms, the same feeds, and the same cultural weight.

The Question of Origin

The photographic industry is already responding to the provenance challenge. The Coalition for Content Provenance and Authenticity (C2PA)³ has developed an open standard for embedding verifiable origin information directly into digital files. A content credential: who made it, when, with what tool, and what changes were made after capture. Adobe, Google, and Sony are among the organisations that have adopted it⁴. In 2023, Leica became the first camera manufacturer to build it into a production camera⁵. The M11-P ships with content credentials embedded at the moment of capture, before the image reaches any editing tool or platform. It is an acknowledgement, from inside the photographic industry, that origin is now a question that requires a technical answer.

Zero Baseline approaches the same question from a different direction. The platform does not yet issue digital certificates. But it considers the task a worthy one. What would it mean to establish a verified digital original for the foundational images of photography’s history? Not a reproduction of the original. A certified primary digital source. Verified by archival record. Locked as the point of origin for every duplication that follows. Because for most of the canonical images of photographic history, nobody currently knows with certainty where the digital chain actually begins. The physical original exists. The first digital file derived from it is rarely documented with the same rigour.

NASA suggests what this could look like at its best: a model where the provenance of every image is open, structured, and built into the process from the start.

Not Good, Not Bad

I do not think the question of AI and the evolution of the image has a moral answer. But it certainly has consequences. We have evolved to see, to form understanding from what we see. The images that shaped that understanding over the last two centuries came from a medium with physical constraints: a box, a lens, a surface sensitive to light. Those constraints produced a visual language. That language is now being learned, replicated, and, if the research is right, slowly degraded by a system that has no physical constraints at all.

The question is not whether this is happening. The question is whether we understand what is being lost when the root of the inheritance is no longer visible.

SOURCES:

Shumailov, Ilia, et al. “AI Models Collapse When Trained on Recursively Generated Data.” Nature, 2024. nature.com/articles/s41586-024-07566-y

Event Horizon Telescope Collaboration. “First M87 Event Horizon Telescope Results.” The Astrophysical Journal Letters, 2019. eventhorizontelescope.org

Coalition for Content Provenance and Authenticity (C2PA). Content Credentials: Technical Standard. 2021–ongoing. c2pa.org

Adobe Content Authenticity Initiative. “How It Works.” contentauthenticity.org/how-it-works

Leica Camera AG. “Leica M11-P: World’s First Camera with Content Credentials Built-In.” Press release, 2023. leica-camera.com/en/photography/content-credentials

NASA. Scientific Image Archive and Open Data. nasa.gov/images

Peter Galison, dir. The Edge of All We Know. 2021. theedgeofallweknow.film

FOOTNOTES:

¹ Shumailov, Ilia et al., AI Models Collapse When Trained on Recursively Generated Data. Nature, 2024.

² Ibid.

³ Coalition for Content Provenance and Authenticity (C2PA). Content Credentials: Technical Standard. 2021–ongoing.

⁴ Adobe Content Authenticity Initiative; Google; Sony. Adoption of C2PA standard. See: contentauthenticity.org / c2pa.org

⁵ Leica Camera AG. “Leica M11-P: World’s First Camera with Content Credentials Built-In.” Press release, 2023.

RESOURCES:

Explore the collection chronologically

Explore by thematic influence

Read the technical genealogy

© 2026 Patricia von Ah — Zero Baseline of Photography, SEETHINK Lab. All rights reserved