LAION: Image Data, AI, and Dispossession

Abstract

This thesis explores the utility of the concept recursive dispossession to analyse infrastructural processes behind generative AI image models. LAION, a non-profit organization which curates datasets for AI training, is chosen as a focal point in the analysis. The dispossession of people from their data is framed as an extension of data colonialism which is empowered by the platformization of the web. This thesis utilizes an infrastructural inversion method, combined with a discourse network analysis to argue that a particular form of dispossession is taking place which manifests itself most saliently through LAION datasets. The analysis shows that dispossession does not take place through one party’s action but through the infrastructural process. By utilizing image data in an unforeseen manner, for generative AI training, the data has gained a new proprietary status. Data is turned into property and by doing so, control is taken from its originators. Legal consent to this process is retroactively assumed. In this process the data is dispossessed in a recursive manner. Because of current legal procedures, there is no existing framework to protect people from their data being appropriated in this way. Additionally, Big tech companies, non-commercial organizations, and academic research groups that work with the data are able to avoid responsibility for the (copyrighted) content of their datasets. These organizations are protected by AI models’ technological constitution which makes it extremely difficult to track or remove data. Currently, the responsibility to protect data lies with the dispossessed. While this dispossession is new in its particularity within AI, it is a continuation of the logic of accumulation on which capitalist expansion is build. As a whole, this thesis emphasizes the need for decolonial alternatives to current digital infrastructures and AI development.