Life cycle assessment of a generative artificial intelligence system: the case of a radio application

Abstract Purpose Artificial intelligence (AI), and particularly generative AI (GenAI), has expanded rapidly in recent years. Although its environmental impacts might be significant, they remain poorly understood and the application of life cycle assessment (LCA) to AI technologies remains in its early stages. The aim of this study is to quantify the life cycle climate impact of an AI-generated radio station named KSPR, marking the first LCA of a system composed of a variety of AI models. Methods In this study, LCA is applied to assess the climate impact of KSPR, which is a research prototype of an AI-generated autonomous radio application developed at KTH Royal Institute of Technology in Stockholm. The functional unit was defined as one hour of generating, broadcasting and streaming radio content. Data was obtained from real-time energy measurements, publications by AI model developers and scientific literature. Sensitivity and scenario analyses were applied to test the reliability of the results and to assess parameter as well as model uncertainties. Results and discussion In the baseline scenario, computer hardware production is identified as the main contributor to KSPR’s life cycle climate change impact (at about 70%), followed by the use of the text-to-speech conversion model Tortoise TTS (about 10%) and streaming (about 9%). Within the computer hardware category, the screen and graphical processing unit contributed significantly. Sensitivity and scenario analysis revealed significant variability in total results and in the hotspots identified depending on assumptions regarding e.g. computer lifetime, number of users of AI models and training parameters. Conclusions and recommendations To the knowledge of the authors, this study provides the first LCA of a composite GenAI system. The findings from this study demonstrate that e.g. computer hardware and inference with Tortoise TTS are key contributors to GenAI applications such as KSPR, while the magnitude of their impacts is highly sensitive to assumptions and data availability. Strategic choices in AI model development, e.g. by reconsidering training practices, and extending hardware lifetimes can reduce impacts. The results emphasise the importance of improved data transparency on AI development for e.g. hardware, training resources, model lifetimes and usage statistics, to enable more accurate and comparable LCAs. This is also important in the context of the LCA methodology applied, such as for allocating impacts of AI model training to specific AI services. Continued advancement of LCA of AI is critical for developing a more complete understanding of the sustainability of AI systems.