Generative AI in the Studio: How Artists Are Actually Using Machine Learning Tools

The conversation around AI and art tends to collapse quickly into a single question — is it real art or not — which skips past a more interesting one: what are artists who actually work with these tools doing with them, day to day, beyond typing prompts into a text box? A small but established group of artists has been building with machine learning since well before text-to-image generators became a mainstream curiosity, and their practices look less like prompting and more like training, curating and physically collaborating with systems they built or heavily customised themselves.

Training Models on Purpose-Built Data

Refik Anadol's Unsupervised, shown at New York's Museum of Modern Art from late 2022 into 2023, is one of the clearest examples of this approach. Rather than using an off-the-shelf generator, Anadol's studio trained a machine learning model on metadata and imagery from MoMA's own public collection, then let the system generate a continuously shifting, large-scale digital work responding in real time to conditions like the weather outside the building. The dataset was not incidental; it was the material, chosen specifically so the resulting piece was, in a sense, MoMA's own collection dreaming about itself, rather than a generic model producing generic imagery.

The Machine as a Drawing Partner, Not a Tool

Sougwen Chung takes a different route to a similar depth of engagement. Chung trains robotic arms on video and sensor data of her own drawing gestures, then performs live alongside the resulting machine, each responding to the other's marks in real time. The work sits somewhere between performance, sculpture and machine learning research, and depends entirely on the system being trained on Chung's own body of movement rather than a general-purpose dataset. It treats the machine less as an image generator than as an improvising collaborator with its own, learned version of the artist's hand.

Making the Dataset the Subject

Other artists use machine learning specifically to interrogate its own limitations. Anna Ridler's Mosaic Virus project involved hand-photographing and hand-labelling thousands of images of tulips to train a generative model, a deliberately slow, manual process that stands in contrast to the scraped, uncredited datasets typically used to train commercial systems. The resulting work draws an explicit parallel between tulip mania in seventeenth-century Holland and speculative bubbles in cryptocurrency and NFT markets, using the labour-intensive dataset itself as part of the piece's argument about value, provenance and where data actually comes from.

The Unresolved Question of Consent

None of this settles the field's most contentious issue: large commercial image generators are typically trained on enormous volumes of images scraped from the internet without the consent or compensation of the artists whose work ends up in that data, prompting ongoing lawsuits and licensing disputes across the industry. Artists working the way Anadol, Chung and Ridler do have effectively sidestepped that problem by building or curating their own training data. It is a slower, more expensive way to work with machine learning, but it is also the clearest evidence that "AI art" is not one practice but several, with very different relationships to authorship, labour and consent.