Every Iris
The Every Iris series presents the instances of “iris” in the LAION-5B dataset in text and image as a series of large scale drawings and matching video work: each half reflecting the same slice of data, The artist has divided the captions into broad categories - flowers, eyes, and people - reflecting the multiple meanings collapsed under a single word within machine-learning systems. Each piece in the series corresponds to a “dip” of approximately two thousand captions from each category. The individual titles (Every Iris: Percentage Monet 7.40%, Every Iris: Percentage Van Gogh 9.54%, Every Iris: Percentage Van Gogh 4.35%) refers to the percentage distribution of items within that slice. The internet (the LAOIN-5B is made up of text to image pairings from the Common Crawl) has always been overwhelmed by ‘slop’, and by looking through each caption it is possible to see how much of this is driven by commerce - images of flowers turn out to be reproductions of Vincent van Gogh paintings printed on coffee cups, shower curtains, or umbrellas.
Process and Research
This work extends the online project Every Single Iris on the Internet and the Synthetic Iris Dataset, whilst also operating as an inversion of Myriad (Tulips). Where Myriad foregrounded the labour of hand-labelling images to make visible the human work underpinning machine learning, this piece refracts that labour through the overwhelming density, repetition, and degradation of networked information. It also is part of my thinking through the Internet works: how it is becoming increasingly unstable and ephemeral, how links break, images disappear and platforms collapse. It is shaped by the condition of a disappearing Internet, that is soon going to be one where there is more machine generated content than human generated - one that machine-learning systems continue to train on even as it erodes. Handwriting is central to the project as both method and metaphor. Writing by hand is one of the oldest recording technologies, requiring time, physical presence, and sustained attention. Set against the scale and speed of automated data extraction, it becomes a deliberate countermeasure to digital decay. By transcribing captions onto paper, the work attempts to hold onto fragments of an Internet that is otherwise slipping out of reach, preserving what would normally remain abstracted, compressed, or lost. The scale of the drawings deliberately references eighteenth- and nineteenth-century scientific atlases, a format that recurs throughout my practice and places contemporary datasets within a longer history of classification and knowledge production. In doing so, it tries to place present-day machine learning within a lineage of earlier efforts to catalogue the world, while acknowledging both the ambition and the fragility inherent in such attempts.
