SWIM (2024)
A meditation on training data, memory, and archives.
The image above is a video, mouseover to play.
SWIM is a meditation on visual archives, the datafication of memory, and training data in the age of generative artificial intelligence. In a reproduction of the AI training process, a digital archival film is slowly diffused as its subject frolics underwater. The actress dissolves into 'training data,' amidst a backdrop made from visualizations of generative AI searching for patterns in the absence of any training data at all. The film is part of a series of works which aims to draw our attention to the content of visual archives that create synthetic images, and what it means to "diffuse" visual culture into the vast latent space of artificial intelligence models.
SWIM was a means of thinking about our cultural archives of visual imagery, and their absorption and re-animation through generative AI systems. There has been a trend to call these archives “datasets,” which erases the cultural context in which they were created. But these datasets are images, films, and records — drawing from a massive trove of content, a “diffusion” of history into novel images constrained by statistics and patterns.
SWIM is a result of "diving” into an archive — The Prelinger Archive, hosted on the Internet Archive — and looking for public domain material to place into dialogue with the products of AI systems. In “SWIM,” a dancer — Nini Shipley, in a film reel from 1952 — swims amongst a layer of noisy, abstract imagery generated from a glitched AI image generation system. By asking AI image generators to create images of noise, they fail, because they are trained to remove noise from images: the system adds noise, and then removes noise, resulting in abstract imagery that does not connect to its training data. The result is a visualization of generative AI as primordial ooze: no data from which to create images, but a constant churning search for patterns.
Shipley’s performance — apparently an “erotic film” according to the Chicago Film Archives — centers the body in ways that are both joyous and leering. Shipley performs with grace, the camera gazes. In the meantime, as is the fate of every image in our archives, the film real — the residue of a body and light on film stock — is dissolved into data, pixels to be arranged as clusters into new patterns. Over time, the images on the screen are overwhelmed by digital noise: Shipley’s conversion to the collective “diffusion” of archival imagery marked as complete.
Diffusion breaks images apart, dissecting them into generalizations, a kind of churning of visual culture. I like to imagine Shipley swimming through this diffusion, not subject to it, but dancing within it, her body imprinting shapes into the patterns that will emerge in the pool and on film and in the contents of future images that rely on her movements to predict future images.
Images in an archive are signposts, aesthetic structures and representations that collectively haunt the images generative AI models produce. This haunting is evident across all AI, when data from the past reaches into the present to shape decisions about the future. I don’t know if I envy Shipley. I see her as a ghost, I see this film as a haunting. I wanted to think through what it feels like to be a digital ghost, our physical likeness turned into pure information and diffused into the collective corpus, forever steering the decisions of machines, like some electric poltergeist.
There’s also something aspirational about swimming in the noise, though: a body immersed in this information debris, but from a remove. It’s this tension between the archives and the models that I hoped to explore with SWIM. The figure of a swimmer’s underwater ballet is slowed down, along with the original score (in the style of Fisher’s sonic hauntology), creating a haunting audio-visual piece that I hope inspires thinking about the entanglements of data analytics, generative artificial intelligence, cultural archives, and visual culture.
