How AI Sees Art

Vackground, a blue and pink abstract background with wavy lines, 2021

Solen Feyissa, pink and blue abstract painting, 2021

As galleries and artists flood the internet with high-resolution images, a silent revolution is underway. The machine is watching - and learning.

I studied photographic art in the days of film and darkrooms, when making an image meant getting your hands dirty. Light was something you handled slowly, not just a technical setting but a tactile presence. There was nothing abstract about it - no metadata, no compression rates, no question about where the image would go once it dried. You made a photograph. You framed it. And that was that. 

Today, photographing an artwork serves a very different function. It’s no longer just about preservation or visibility. In many cases, it’s feeding systems we can’t see, fuelling uses we never imagined, and contributing to datasets that have nothing to do with art history or documentation. 

What once lived on contact sheets and slides now lives in the hands of algorithms. 

And the truth is: they’re learning from us. 

Cash Macanaya, Hold My Hand, 2023

Hugol Halpingston

The New Apprenticeship

Generative AI models like Stable Diffusion, DALL·E, and Midjourney are built on enormous datasets -billions of image-text pairs scraped from the open internet. These datasets include artworks. Paintings. Installations. Photographs of photographs. Museum holdings. Gallery archives. Artist websites. Anything accessible by a web crawler is fair game, unless explicitly blocked - and most art institutions haven’t had the need (until now) to erect those kinds of digital defenses.

The Aesthetic Feedback Loop

For galleries and artists, this creates a strange paradox. High-quality photographs are essential: for press, for collectors, for credibility. But those same images, shared widely and with pride, may end up contributing to datasets that allow algorithms to mimic the very aesthetics they document. This isn’t theoretical. Artists are already seeing it. A growing number of practitioners have reported AI- generated works that closely resemble their signature styles - colour palettes, brush strokes, compositional rhythms - without attribution, permission, or even acknowledgment. And the source material? Most likely, the images they or their galleries uploaded in good faith, to promote their work.

From Representation to Instruction

Traditionally, photographing an artwork was an act of translation - a way to render the physical into the digital. In the current climate, it's something else entirely: instruction. The machine doesn’t just see the image - it learns from it. Learns how to simulate oil paint, replicate lighting, recreate style. And because generative AI doesn't copy images directly, but creates new versions from statistical patterns, it's nearly impossible to trace the original influence. The mimicry is plausible, but deniable.

This raises difficult questions for the art world:

- What happens when an AI-generated pastiche becomes more shareable than the original?

- What rights do artists retain when the style is taken, but not the work?

- And how much responsibility falls on the simple act of uploading a high-res image?

Legal Loopholes and Ethical Grey Zones

There are no clear protections - yet. Copyright laws weren’t designed for this kind of machine learning. Getty Images is currently suing Stability AI for allegedly copying 12 million of its assets. But for independent artists and smaller galleries, enforcement is elusive, expensive, and largely retrospective.

Some institutions are beginning to take defensive steps: blocking crawlers with robots.txt files, adding anti-AI meta tags to their websites, and limiting image resolutions online. Others are exploring digital watermarking or licensing agreements that explicitly prohibit AI training use. (See our guide: Protecting Artwork Images in the Age of AI.)

Still, the infrastructure to protect visual material is playing catch-up with the tools that extract it.

It isn’t just stock photos or casual

images being absorbed.

Fine art, museum archives, and gallery photography are all part of the mix.