Artificial Intelligence’s Environmental Impacts:  Exploring Research on “Red AI"

It is a penetrating glimpse into the obvious to say that the development and use of artificial intelligence is growing by leaps and bounds. The global market for AI in 2024 was US$136 billion and is forecast to reach $668 billion by 2030 [1]. As with many technologies, there is potential for both environmental benefit and environmental harm. And, as with any technology, some impacts will be direct—e.g., reductions or increases in carbon emissions—and some will be indirect—changes in behavior, business practices, markets, and institutions that influence environmental outcomes.

While there is vast public discussion about how AI will affect markets, jobs, media, and many other parts of our lives, the discussion of the environmental implications is not nearly as extensive. Until very recently, much of the discussion of artificial intelligence’s environmental impacts had been limited to speculation about how AI might be used to fight climate change and improve environmental management. Discussion and research on the negative environmental implications of AI are growing but are not nearly as developed. The detrimental implications of AI, however, can be as important as the beneficial effects. Borrowing from a seminal 2020 paper by Schwartz and colleagues, we label the detrimental environmental and energy impacts resulting from the development and use of AI—and from changes in production and consumption that arise from that use—as “Red AI.”

Given the likely ubiquity of AI in the coming years, an understanding of how the research literature on AI and the environment has evolved is useful. This paper uses bibliometrics—the quantitative investigation of the pattern of research publications using publication counts, citations, and content analysis techniques—for this purpose. The analysis presented here is a proof of concept showing the benefits, challenges, and lessons from the use of bibliometrics and related techniques to understand the development of research on Red AI. This paper describes the first step in this process: identifying the relevant research; a subsequent paper will present the results of the quantitative examination of patterns and trends in the compiled research publications.

Identifying and locating research on Red AI is a challenge for several reasons. Many potential impacts of AI revolve around energy consumption and carbon emissions from its use with other technologies, making the attribution of environmental impact ambiguous. If a drone or robot uses AI, is the energy consumption from those devices attributable to AI? In contrast, the topic of energy consumption from “AI compute” can be deemed a topic that falls comfortably in the scope of this analysis. The conceptual challenges of defining Red AI are complemented by an operational impediment. In searching bibliographic databases, it is extremely difficult to construct search strategies that differentiate between research on environmentally beneficial AI and potentially harmful AI. No automated processes used in the work presented in this paper—Boolean searches in citation indices, natural language processing of titles and abstracts, and other more elaborate forms of analysis—successfully accomplished the desired differentiation. The different venues for publication of computer science research and other types of research compound the difficulty of investigating patterns of publication—computer scientists typically publish in conference proceedings, whereas environmental and many other disciplines prioritize publishing in academic journals. Conference proceedings are indexed in citation indices, but not completely, especially in Web of Science, the source on which bibliometricians often rely.

We identified 594 research publications on Red AI published as of March 2024. Research publications in computer science constituted a large proportion of the papers we identified in part because we included both research on the impacts of AI and research on possible remedies to those impacts. Our preliminary analysis suggested that crossover between computer science and non-computer science literature is limited. While a moderate number of papers quantify energy or carbon impacts of AI training and inference, very few include impacts across the entire product life cycle or scale impacts or potential improvements arising from widespread deployment.

Consequently, the dataset of papers created in the processes described in this paper is illustrative of, rather than a comprehensive compilation of, Red AI research. The paper describes the lessons learned and unexpected challenges in identifying a dataset of 594 research publications on Red AI. Because the research literature compiled in this study encompasses papers published up through March 2024, it only partially reflects the significant increase in attention to Red AI that developed after ChatGPT and other generative AI models were made publicly available.

Based on the analysis presented in this paper, we recommend careful attention to the scope of research in constructing datasets for bibliometric analysis. The types of environmental impact, and the types and uses of AI included, and the sources of publication data, will strongly shape any conclusions about the character and evolution of Red AI research. Because of the limitations of keyword-based searching we encountered, we also emphasize the use of snowball searching—sequentially using the reference lists of papers, or the citations to papers, to identify additional related papers—as a necessary component of searching the research literature related to Red AI.