"Erosions” is a body of experimental video works created through deliberate mis-training: I fed a CycleGAN two datasets that shouldn’t converge—selfie videos and sea-horizon recordings. The system began to hesitate. Faces leak into water. Ghostly almost-figures surface—not enough to name, enough to feel.
The work resists the dominant digital aesthetics of hyper-clarity, control, and instant recognizability. Instead, it stages a fragile system in which instability becomes method: low resolution slows the eye, blur extends time, and confusion enters as texture. The “I” erodes—no longer singular, but produced between AI translation, my dataset, and the viewer’s attempt to recognise a face. Trained exclusively on my own archive, the work refuses mass-scraped datasets and their generic output, insisting instead on situated data and accountable aesthetics. It becomes a plea for agency over data—without surrendering to opaque infrastructures and statistical averages.
