Environmental Impact Assessment of Large Language Models Beyond Carbon Footprint

Large Language Models (LLMs) are AI models used to elaborate and generate text. LLMs are particularly costly to the environment due to their large size and use on a large scale. Rapid advances in research have made large language models ubiquitous, to the point where their use is being proposed as a solution to the challenges of climate change. In light of this, the assessment of their environmental impact prior to their application is crucial. The purpose of this thesis is to investigate the branch of research that assesses the environmental footprint of large language models. This is done by examining, first, what features of these models contribute to their impact and what are their consequences, and second, by reviewing empirical studies that have assessed the environmental footprint of LLMs. As a result, the complex infrastructure on which LLMs rely contributes significantly to the overall impact of the models. Furthermore, their use facilitates climate disinformation and has consequences on workers and marginalized communities. This research also provides a taxonomy for assessing the environmental impacts of LLMs. Namely, the taxonomy consists of (1) the unit of impact measurement, (2) the contribution of each life cycle phase, and (3) the scope of impact. Using the taxonomy to evaluate the selected studies highlights the limitations of current research, which is overly focused on carbon emissions and direct impacts. This leaves out important aspects such as the hardware life cycle assessment. The lack of a standardized approach makes it difficult to comprehensively assess the environmental impact of LLMs. Further work is needed to achieve this goal, and this research contributes by defining some directions for this discussion. Effective solutions must go beyond the concept of carbon efficiency, recognize the responsible actors, and challenge the paradigm of unbridled growth that caused the ecological crisis in the first place.