AI as a regional product - Adversarial Design for Public Debate on LLMs and Deep Culture AI as a regional product

"Bigger is better" has long been the adage driving the development of large AI systems: the more data, the better the performance. However, a shift is emerging. Generative AI systems, which are increasingly trained on data produced by earlier iterations of similar models, appear to be declining in quality. This phenomenon-referred to as model collapse-may lead to a broader knowledge collapse, in which growing dependence on generative AI, such as large language models, results in erosion of knowledge quality. Furthermore, large AI systems could be contributing to cultural homogenization, as globally trained models tend to obscure local and contextual differences. Though this topic is gaining traction in scholarly research, the audience at large is hardly involved in the discussions. The research-through-design project AI as a Regional Product, in which a model was trained on region-specific visual data, provides a tangible framework for fostering a public discussion on the impact of Large Language Models on culture and vice versa.