Every Single Iris on the Internet (the first 100,000 images)

The internet is the secret foundation of machine learning, as machine learning cannot work without a dataset and, for the most part, datasets come from the internet. Because this type of technology - machine learning, artificial intelligence - is changing so rapidly there can be an assumption that the data that trains it is also changing quickly. But that is not really the case. Once produced, they are very rarely reviewed or updated. Before LLMs and other large models, datasets that were commonly used in this area were sometimes 10 or 15 years old; now, with the type of machine learning that requires millions if not billions of inputs, the entire internet is being sucked up for it to learn from. But it is sometimes an internet that no longer exists: the common crawl, which scraps images and their accompanying captions, has been running and has data for seventeen years; LAION (a dataset that is often used for text to image training and which this project is based off) contains pairings from a full decade ago. Because of this a number of these pairings are broken - the image or description no longer exists and the only trace or ghost of it is left in the index of this massive database. The world that exists in this internet is not the world we live in today. This work runs through these images, which have been sorted by me, showing each flower until it hits open a broken link or ghost iris when it stops. It is virtually impossible to see every single thing that has been used to make them but I have been obsessed with trying to do so. There is something about actually looking at each of these images with a human eye - rather than just being a pattern that is understood from statistical significance - that speaks to a different type of engagement with the information, one of looking and seeing and paying attention rather than being unknowingly ground up behind the scenes. If a viewer stays long enough to the end of the work a different type of iris appears - one that has been trained on synthetic copies of irises. Part of the reason why there is resistance to using the internet as it is now is because it is riddled with synthetic images, created by AI, and there is sometimes a cut off of using data scrapped from after Dalle and Chatgpt became popular, the web becoming too noisy and messy to be of use. These irises are a different type of ghost - uncanny and unreal versions of the thing in question, echoes of echoes—no longer grounded in the physical world but in the recursive logic of machine learning.