Life Cycle Assessment of Artificial Intelligence Applications: Research Gaps and Opportunities

Applications based on Artificial Intelligence (AI) are increasingly being adopted, especially since tools like ChatGPT or Dall-E have become available to the public and have gained popularity. While its benefits are manifold, AI also has a significant environmental impact: AI technology usually relies on hardware which consumes a significant amount of energy, such as GPU servers. Furthermore, users normally access AI-based services via their own end devices, which also consume energy. The energy side of training and using AI models can be roughly estimated today, for example using libraries like CodeCarbon. However, a realistic estimate of AI’s environmental impact needs to include not only the energy consumption from the use of the AI, but also factors such as the (critical) raw materials needed for the construction of the hardware, and the consumption of materials and energy for the maintenance and use of the hardware throughout its life cycle. This is not feasible today. Systematic life cycle assessment of AI technology is still in its infancy. It should not be limited to calculating the CO2-equivalent, but take into account different relevant environmental impact indicators from resource use depletion to human toxicity to water consumption. This paper reviews the literature on this topic, identifies gaps and opportunities for research and argues for developing an LCA methodology for AI. It further presents an adaptation of an existing LCA methodology to AI as a first step in this direction.