JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) systems increasingly contribute to global energy consumption, greenhouse gas emissions, and water use, yet these impacts remain inconsistently measured and poorly reported in the research literature. Existing tools primarily focus on operational energy during model training and often omit critical contributors, including hardware lifecycle emissions, cooling overhead, and inference-level impacts. The absence of standardized, accessible, and reproducible environmental accounting frameworks limits comparability across studies and impedes environmental transparency in AI research.

Objective:

This study aims to develop and validate a transparent, lifecycle-informed framework for quantifying and reporting the environmental footprint of AI workloads and to propose a standardized environmental reporting and labeling approach suitable for routine inclusion in AI research publications.

Methods:

We developed the AI Environmental Footprint Calculator, a browser-based, open-access tool that integrates four core components of environmental accounting: (1) operational energy use derived from hardware power and runtime; (2) embodied emissions from hardware manufacturing and end-of-life amortized over device lifetime; (3) cooling and infrastructure overhead using configurable Power Usage Effectiveness (PUE) and optional Water Usage Effectiveness (WUE); and (4) inference-level granularity expressed as carbon dioxide equivalent (CO₂e) per 1,000 tokens or per request. The calculator incorporates regional and time-resolved grid carbon-intensity data and provides both minimal-input and expert modes. Validation was performed using three representative scenarios: a small laboratory fine-tuning task, a mid-sized academic research cluster, and a large-scale industrial training workload. Sensitivity analyses assessed the influence of key parameters, including grid emission factor, PUE, and hardware utilization.

Results:

Across validation scenarios, the calculator produced consistent and interpretable estimates of energy use, CO₂e emissions, and water consumption, while revealing systematic underestimation by operational-only tools. Inclusion of hardware lifecycle and cooling effects increased total emissions estimates by approximately 20%–40%, depending on workload scale. Scenario analyses demonstrated that algorithmic optimization, carbon-aware scheduling, and infrastructure choices can substantially reduce environmental impact. Based on these outputs, we introduce an AI Environmental Label (A–E grade) that benchmarks workloads using normalized carbon-intensity metrics for training and inference, enabling intuitive comparison across studies.

Conclusions:

This work presents a comprehensive, reproducible framework for environmental accounting in AI research that bridges operational measurement with lifecycle assessment and inference-level reporting. By combining methodological rigor with accessibility, the proposed tool and labeling system support standardized environmental disclosure and informed decision-making. Adoption of such frameworks can facilitate transparency, comparability, and responsible scaling of AI systems across research environments.