aiculedssul

This experimental project investigates the phenomenon of model collapse in artificial intelligence through systematic iteration of image generation. The work uses Stable Diffusion, a state-of-the-art deep learning model, to generate images from the prompt "a single chair on a white background". These generated images are then reused as training data for subsequent iterations of the model, creating a closed feedback loop within the system. The documentation spans six generations of this process, revealing the progressive deterioration of the model's representational capabilities. As the cycle continues, the initially clear and figurative representations of the chair undergo a transformation, degrading into abstract digital artefacts. This degradation is an example of model collapse - a significant phenomenon in machine learning, where the output of an AI system becomes increasingly homogeneous and corrupted when trained exclusively on its own generated data. The project acts as a technical demonstration and critical examination of the inherent vulnerability of AI systems to data feedback loops. It highlights how computational knowledge systems can degrade when isolated from diverse, real-world training data, raising fundamental questions about the stability and reliability of machine learning systems. Through the familiar form of a chair, this work materialises the abstract concepts of model collapse and data cannibalisation. The progression from recognisable furniture to digital noise serves as a visible manifestation of the subtle yet profound ways in which artificial intelligence systems can lose their grip on semantic understanding, revealing the fragile nature of computational ontology in closed learning environments.

Credits: Artwork, text, digital production, film: Ègor Kraft AI engineering: Artem Konevskikh Voice: Sara Woodgate.