Understanding the Environmental Impact of Generative AI Services

Abstract

Understanding the Environmental Impact of Generative AI ServicesThe past few decades have been marked by the ever-increasing presence of digital technology.This growth, often called digital transformation, places a heavy burden on our environment.We are now facing a potential new phase of digital transformation, 6 represented by the emergence of generative AI (GenAI), a subfield of artificial intelligence focused on generating new content, such as human-like text, code, and images. 14In particular, the deployment of GenAI as a service, such as ChatGPT or Stable Diffusion, is raising questions around sustainability.The sustainability of any computing technology, however, cannot be addressed without a way to evaluate its environmental impact.And though we hope that some digital applications will help meet sustainability goals by reducing the impact of human activities on the environment, it remains difficult to measure the positive or negative environmental impacts of digital technology. 3,32ere, we present a methodology, along with its application, to assess the environmental impact of Stable Diffusion as an end-to-end service.To better evaluate the sustainability of this GenAI application, we include not only the carbon impact but also the consumption of metals in a life cycle assessment (LCA) approach.We also question the current methods used to estimate the electricity consumption used for training GenAI models.We propose a transparent, reproducible approach based on sampling through experiments, presenting measurement tools and methods that aim to be more reliable and holistic to allow us to rethink sustainability challenges and improvements in the use of GenAI services.Our impact methodology is based on previous work, 2 extending its scope to data storage and all training costs, and differentiating operational emissions from those of the use phase of the hardware life cycle, and embodied emissions from those of other phases.We begin with a review of the current approach to environmental impact in the fields of AI and GenAI.We then present our tool to enhance the current way in which the environmental impact of GenAI is assessed.Finally, we show how our contributions are helping to frame new obstacles and sustainability challenges around GenAI services. Overview and Limits of Environmental Impact Assessment for AIThe rise of generative AI: A quick review.Though the term generative AI was coined in 2014, 15 since the end of 2022 it has gained notoriety, and thus great interest, well beyond AI research.The sudden and significant rise of this term, however, should not conceal the long research history behind the technology. 9Developing GenAI models requires collecting data and learning from it, which includes first selecting the best model structure and learning algorithm for the given task and then applying this algorithm to the model and the data.The first step hides an expensive development process, as GenAI models are usually composed of several already-developed models.The second step is called training.Once a model has reached the targeted quality level through training, it can be used on new data, referred to as the inference phase.