BIRDS: Characterizing and Understanding Biodiversity Impact of Large Language Model Serving

A new framework called BIRDS aims to quantify the biodiversity impact of serving large language models, moving beyond the commonly tracked metrics of carbon and water use to assess ecosystem damage from AI inference workloads [1][2]. The framework, detailed in a paper submitted on 26 May 2026, is formally named the Biodiversity Impact of Request-Driven LLM Serving [1][2]. It defines request-level functional units to measure the ecological consequences of individual queries processed by large language models [1][2]. The methodology accounts for both operational impacts, from energy consumed during inference, and embodied impacts tied to hardware manufacturing and disposal [1][2]. A key component of the framework is a metric called Quality-Normalized Biodiversity Impact, or QNBI, which jointly analyzes ecological harm and the quality of a model's responses [1][2]. The authors report that across diverse workloads, model architectures, GPU types, and geographic regions, biodiversity impact accumulates at scale, revealing tradeoffs that can inform quality-aware serving decisions [1][2]. Forests, which the United Nations Food and Agriculture Organization defines as land spanning more than 0.5 hectares with trees higher than 5 meters and a canopy cover of more than 10 percent, covered approximately 31 percent of the world's land area in 2025 [5]. These ecosystems account for 75 percent of the Earth's gross primary production and contain 80 percent of its plant biomass [5]. The BIRDS framework provides a structure to connect the energy and material demands of AI infrastructure to pressures on such ecosystems, which are already experiencing fragmentation and deforestation driven by commodity production [5]. Almost half of Earth's forest area remains relatively intact, while 9 percent exists in fragments with little or no connectivity, according to global assessments [5]. The paper was posted on arXiv, an open-access repository operated by Cornell University that hosts preprints in fields including quantitative biology [1]. The work appears under the site's quantitative biology section and is associated with arXivLabs, a platform for experimental community-driven projects [1]. The authors did not provide direct quotes in the abstract, but the text states the framework "reveals that biodiversity impact accumulates at scale and exposes actionable quality-aware serving tradeoffs" [1][2]. The research contributes to a growing body of work examining the full environmental footprint of artificial intelligence systems, extending analysis beyond greenhouse gas emissions to include biodiversity-related pathways [2].