关于科学领域人工智能（AI）环境影响的思考

© International Science Council, 2025 To cite this report: International Science Council (2025).Considerations on the environmental impact of AI inscience. DOI: 10.24948/2025.10 Authors:Denisse Albornoz and Natalia NororiReviewers:Marcel Dorsch and Gong KeProject coordination:Dureen Samandar Eweis, Vanessa McBrideProject chair:David Castle Funding acknowledgement:This work was carried out with the aid of a grant from theInternational Development Research Centre (IDRC), Ottawa, Canada. The views expressedherein do not necessarily represent those of IDRC or its Board of Governors. Design:Mr Clinton Cover photo:Frank Ramspott About the International Science Council The ISC is an international non-governmental organization with a unique global membershipthat brings together 250 international scientific unions and associations, national andregional scientific organizations including science academies, research councils, regionalscientific organizations, international federations and societies, and academies of youngscientists and associations. The ISC works at the global level to catalyse change by convening scientific expertise, adviceand influence on issues of major importance to both science and society. KEY TAKEAWAYS4 ABOUT THIS PAPER5 INTRODUCTION6 SECTION 1: FOUNDATIONAL CONCEPTS AND FRAMEWORKS 1.1Green AI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .81.2Sustainable AI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .91.3Ethical and responsible AI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10 SECTION 2: ENVIRONMENTAL IMPACT OF AI ACROSS THE LIFECYCLE 2.1Environmental impacts of AI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .112.2A lifecycle approach to assessing environmental impact. . . . . . . . . . . . . . . . . . . . . . . . . . .12-Software layer12-Hardware layer132.3Estimating direct environmental costs across the AI lifecycle. . . . . . . . . . . . . . . . . . . . . . .13-Estimating operational costs13-Estimating embodied costs142.4Transparency and reporting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14 SECTION 3: STRATEGIES FOR REDUCING DIRECT ENVIRONMENTALCOSTS OF AI APPLICATIONS IN SCIENTIFIC RESEARCH 16 3.1Reducing environmental costs in the software life cycle. . . . . . . . . . . . . . . . . . . . . . . . . . .16-Data management16-Algorithmic optimization17-CASE STUDY 1: InkubaLM: A small language model for low-resource African18languages3.2Reducing environmental costs in the hardware life cycle. . . . . . . . . . . . . . . . . . . . . . . . . .18-Energy-efficient hardware and computing18-CASE STUDY 2: TinyML enables data analysis and research in low-cost devices193.3Energy-efficient data centres. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19-CASE STUDY 3: Green AI infrastructures across regions203.4Decommissioning and end-of-life management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .21 CONCLUSION REFERENCES23 APPENDIX 1: GLOSSARY28 Key takeaways •There is limited awareness and evidence about the environmental costs of usingartificial intelligence (AI) in scientific research.This article offers frameworks andtools scientists and research institutions can consider to assess the environmentalimpacts of their research as part of a more sustainable, ethical and responsible use of AI inscience. •Addressing the environmental impact of AI requires a multi-dimensional approach.Scientists and researchers who are planning to incorporate AI into their workflows need toassess tools in light of their scientific value, social equity and environmental costs acrossthe entire AI life cycle, with attention to rebound effects and long-term consequences. •Adopting more resource-efficient AI models has environmental and social benefits.Smaller, local and frugal approaches to AI can improve accessibility, affordability,transparency and social inclusion around the use of AI, especially in diverse and resource-constrained research contexts. About this paper This paper examines the environmental implications of applying artificial intelligence (AI)in scientific research. It serves as a primer for scientists, research institutions and sciencepolicy-makers who seek to understand various approaches to addressing the environmentalimpact of AI in science. In addition, it offers guidance on how reducing environmentalcosts can contribute to the broader goals of sustainability and ethical AI use in researchenvironments. Although evidence on the specific environmental impacts of AI in scientific research is stillemerging, the paper provides conceptual frameworks and practical tools to help a