Livscykelanalys av Generativ AI : Fallstudie av en radiostation

Generative AI (GenAI) represents a new emerging technology of machine learning models which enables the autonomous generation of content such as text, images and audio at request. While GenAI offers opportunities of innovation and solutions, questions have been raised regarding the environmental impact of GenAI. Important concerns regarding the environmental impacts of GenAI is the high energy consumption and material use related to training and operating AI models and its contribution to climate change. A possible suitable method to assess the impacts associated with GenAI models is Life Cycle Assessment (LCA), a method to quantify the potential environmental impacts of a system throughout its life cycle. LCA could provide valuable insights on how to improve GenAI models. However, currently there exists very few LCA studies on GenAI that assesses the full life cycle of it. One recent application of GenAI has been developed at KTH called Stochastic Pirate Radio (KSPR), an AI generated radio station. The environmental impacts of KSPR were assessed in this study using LCA methodology. The focus of the study was to assess KSPR’s contribution to climate change. Therefore, data on energy consumption was primarily gathered by measurements at KTH (for the use phase) and through literature studies. While aiming at identifying the environmental impacts of KSPR, the aim was also to provide insights on the methodological challenges of conducting LCA of AI. The results of the study showed that the most significant contributions to climate change came from the hardware (computer) needed to operate KSPR, the training and use of the text-to-speech model TorToise TTS and the end-user streaming of KSPR. For the LCA methodology on AI several challenges were identified, including lack of methodological standards and difficulties in data collection for both hardware and software due to scarce data availability and lack of transparency resulting in high uncertainty in results. The restricted information about AI models complicates the data collection, which forces the use of estimations and assumptions that have a significant impact on the reliability of the results. In conclusion, by identifying the environmental hotspots of KSPR recommendations can be made on how to improve it: using pre-trained AI models that are used by many others and using optimized hardware while also trying to extend its lifetime as long as possible. To identify this, not only for KSPR but GenAI systems in general, LCA can be a useful tool, but further methodological developments are needed to increase certainty and reliability.