life cycle assessment

Carbon emissions from generative AI inference have been systematically underestimated due to ambiguous system boundaries, lack of standardized metrics, and methodological biases in existing assessments. This study employs a scoping review to systematically analyze carbon footprint measurement tools, classification frameworks, and research trends across the full training–inference lifecycle. We propose (i) a standardized measurement protocol, (ii) a dynamic evaluation framework, and (iii) a multidimensional sustainability assessment system—novelly integrating user behavior and system boundary definitions into a unified analytical model. Our analysis uncovers nonlinear relationships between carbon emissions and model scale, prompt complexity, and deployment environment, identifying key leverage points for mitigation. The findings establish a reusable methodological benchmark and practical guidelines for AI environmental governance across technical, social, and operational dimensions.

Current environmental impact assessments of the AI lifecycle suffer from tool heterogeneity, insufficient coverage of water usage and value-chain stages, and poor cross-study comparability. To address these gaps, this paper proposes a unified operational definition of Green AI and introduces a five-stage lifecycle framework—encompassing hardware manufacturing, model development, training, deployment, and reuse. It explicitly distinguishes Green AI (energy-efficiency–focused) from Sustainable AI (holistically sustainable across environmental, social, and economic dimensions). Methodologically, the framework integrates Life Cycle Assessment (LCA), PDCA-based governance, edge–cloud co-optimized hardware strategies, and a calibrated multi-level metrics system combining estimation and empirical measurement. The resulting methodology enables vendor-agnostic, reproducible quantification of energy consumption, carbon emissions, water use, and embodied impacts. This significantly enhances transparency and comparability, providing researchers, engineers, and policymakers with an evidence-driven, actionable guide for sustainable AI development and deployment.

Widely deployed AI-assisted carbon footprint calculation systems lack standardized, credible evaluation criteria; existing guidelines are outdated, benchmark datasets are scarce, and uncertainty analysis remains infeasible at scale. Method: We propose the first comprehensive credibility verification framework specifically designed for AI-assisted carbon accounting systems. Departing from conventional itemized auditing, our system-level approach integrates three core metric categories: benchmark testing, data quality indicators, and uncertainty characterization—tailored to use cases such as corporate GHG accounting and product-level hot-spot identification. The framework was developed via iterative demand analysis, standards drafting, and empirical piloting, incorporating life cycle assessment modeling, AI-to-domain mapping techniques, and statistical uncertainty quantification. Contribution/Results: It enables automated, high-fidelity credibility assessment with scalability, reproducibility, and verifiability—serving practitioners, third-party auditors, and standardization bodies.

Current Green AI research lacks integration with established quality models and service-level agreements (SLAs) from the ICT domain, hindering automated analysis, comparison, and certification of machine learning model energy consumption. Method: We propose a sustainability quality model that unifies Green AI principles with standardized model reporting frameworks (e.g., Model Cards) and design a domain-specific language (DSL) for formal modeling of energy-related characteristics. This enables structured specification of energy metrics, cross-model automated benchmarking, and verifiable sustainability compliance assessment. Contribution/Results: First, we systematically introduce quality model theory into Green AI evaluation. Second, we extend Model Cards into an extensible, executable format supporting rigorous energy-aware documentation. Third, we establish a practical technical foundation for energy efficiency transparency, model selection optimization, and green certification of AI systems—bridging the gap between sustainability goals and operational ICT governance.

This study addresses the widespread lack of rigor in model construction, calibration, and integration within current ICT life cycle assessment (LCA) practices, which undermines result credibility. To tackle this issue, the work proposes four foundational trustworthiness criteria for ICT LCA—model provenance, clearly defined scope, end-to-end traceability, and obsolescence management—and introduces an open computational LCA framework grounded in these principles. The framework integrates dependency graph modeling, a versioned model repository, automated enforcement of completeness constraints, and a structured model taxonomy. It effectively uncovers common misuses and structural flaws in existing LCA studies, thereby offering a viable pathway toward transparent, reproducible, and sustainably evolvable assessments of ICT environmental impacts.

Rapid AI advancement has exacerbated challenges in assessing environmental impacts and lacks transparency in corporate carbon footprints. Method: This study develops an enterprise-level AI environmental impact assessment framework covering the full lifecycle—training, inference, hardware manufacturing, and end-of-life—enabling non-LCA experts to quantify carbon footprints. It introduces a novel multidimensional projection model integrating scaling laws, semiconductor energy-efficiency trends, and grid decarbonization pathways, alongside the first “Environmental Return on Emissions” (EROE) metric to align AI development with net-zero goals. Results: Combining streamlined LCA, IPCC electricity scenarios, empirical compute analysis, and generative AI adoption forecasts, we find: (1) large language models consume 4,600× more energy than traditional models; (2) under high-adoption scenarios, global AI electricity demand will increase 24.4× by 2030; and (3) isolated technological optimizations are insufficient to meet climate targets.

Machine learning (ML) sustainability encompasses environmental, social, and economic dimensions, yet current practice overemphasizes carbon footprint while neglecting holistic assessment and implementation barriers. Method: This study conducts the first integrated investigation—combining semi-structured interviews (N=32) with a large-scale online survey (N=203)—to systematically examine ML engineers’ perceptions, gaps, and constraints regarding triple-bottom-line sustainability. Contribution/Results: Findings reveal widespread awareness deficits among practitioners, compounded by the absence of standardized sustainability metrics, engineering tooling, and institutional support. We identify critical organizational and infrastructural bottlenecks hindering sustainable ML adoption. Accordingly, we propose three actionable pathways: (1) development of structured, role-specific engineering guidelines; (2) establishment of a multi-dimensional sustainability measurement framework; and (3) advancement of cross-industry policy coordination. This work provides the first empirically grounded, cross-dimensional analysis of ML sustainability challenges and offers concrete, implementable recommendations for operationalizing sustainable ML engineering.

This study addresses the lack of standardized methodologies for Scope 3 Category 1 carbon accounting in enterprise AI inference services, where current practices often overestimate emissions by 10–40× due to reliance on coarse-grained industry-average emission factors. The authors propose the first four-tier hierarchical accounting framework aligned with CSRD compliance requirements, which degrades systematically—from a token-level physical energy consumption model to expenditure-driven environmentally extended input-output (EEIO) analysis—based on data availability. Integrating GPU energy benchmarks (ML.ENERGY v3), regional grid carbon intensities (EPA eGRID, Ember), and water efficiency metrics, the framework reveals a previously undocumented carbon-water trade-off: low-carbon regions may entail high water consumption, influencing data center siting decisions. An empirical assessment of a 200-person European firm shows annual AI inference emissions under 1 ton CO₂e, indicating that regulatory challenges stem from methodological gaps rather than emission magnitude, while also uncovering water-carbon synergies overlooked by mainstream ESG tools.

Scope 3 greenhouse gas emissions are challenging to analyze accurately due to sparse disclosure, heterogeneous formats, and a lack of traceable evidence. This work proposes the first dual-granularity framework for extracting Scope 3 emissions data at both organizational and building levels, featuring full traceability of supporting evidence. By integrating optical character recognition (OCR), large language models (LLMs), rule-based engines, and table reconstruction techniques, the framework enables end-to-end extraction of high-precision, interpretable Scope 1–3 emissions data from real-world ESG reports. Concurrently, the study introduces the first multimodal, evidence-annotated dataset designed to support reliable integration and transparent provenance tracking of heterogeneous emissions disclosures.

Carbon life cycle assessment (LCA) suffers from labor-intensive, costly, and time-consuming manual process modeling. Method: This study proposes the first LLM-based automated process generation framework for LCA, integrating the ISO 14040/44-compliant LCA classification system with the world knowledge of large language models (LLMs), leveraging prompt engineering and structured process modeling. Accuracy is quantified via F1-score against ground-truth processes extracted from real LCA documentation. Contribution/Results: Evaluated on 10 diverse case studies, the framework achieves a 62% F1-score; most outputs are either fully correct or contain only minor deviations. Each analysis costs under USD 1 and completes in under 10 minutes. Compared to conventional expert-driven approaches and chain-of-thought prompting, this method substantially reduces human effort while advancing accuracy, computational efficiency, and accessibility—establishing a scalable new paradigm for rapid carbon footprint assessment of consumer products.

This study addresses a critical gap in current AI efficiency evaluations, which typically focus only on isolated training or inference phases and fail to capture the full lifecycle resource consumption and environmental impact of AI systems. To overcome this limitation, the work introduces, for the first time, a comprehensive Life Cycle Assessment (LCA) framework tailored to machine learning. This approach systematically integrates energy use and embedded environmental costs across all stages—including hardware manufacturing, model training, and deployment—thereby transcending the narrow scope of conventional assessments. By providing a holistic and accurate methodology for evaluating sustainability, the proposed framework offers researchers, developers, and policymakers a robust tool to guide more environmentally responsible design, deployment, and regulation of AI technologies.

This work addresses the lack of fine-grained and accurate carbon emission measurement methods during large language model (LLM) inference, which hinders informed sustainability decisions. To bridge this gap, the paper introduces the first systematic reference framework to guide the design of carbon estimation tools for LLM inference and presents SEAL—an early implementation enabling per-prompt carbon footprint assessment. SEAL integrates multi-benchmark-driven modeling, fine-grained energy consumption mapping, and a dedicated carbon estimation algorithm, substantially improving both accuracy and generalizability. Preliminary experiments demonstrate its effectiveness, laying the groundwork for standardized, reproducible sustainability evaluation within the LLM ecosystem.