更智能、更精简、更强大：资源高效型生成式人工智能与数字化转型的未来

Cite as: UNESCO (2025).“Smarter, Smaller, Stronger: Resource-Efficient Generative AI & theFuture of Digital Transformation”Published in 2025 by the United Nations Educational, Scientific and Cultural Organization(UNESCO), 7, Place de Fontenoy, 75007 Paris, France.© UNESCO 2025CI/DIT/2025/ER/01This study is available in Open Access under the Attribution-ShareAlike 3.0 IGO (CC-BY-SA 3.0IGO) license(https://creativecommons.org/licenses/by-sa/3.0/igo/).By using the content ofthis study, the users accept to be bound by the terms of use of the UNESCO Open AccessRepository (https://www.unesco.org/en/open-access/cc-sa).The designations employed and the presentation of material throughout this study do notimply the expression of any opinion whatsoever on the part of UNESCO concerning the legalstatus of any country, territory, city or area or of its authorities, or concerning the delimitationof its frontiers or boundaries.The ideas and opinions expressed in this study are those of the authors; they are notnecessarily those of UNESCO and do not commit the Organization.Authored by: Leona Verdadero1*, Ivana Drobnjak2*, Hristijan Bosilkovski2*,Zekun Wu23,Emma Fischer,and María Pérez-Ortiz2.1UNESCO2University College London3Holistic AI* First three authors contributed equal work.Printed by: UNESCOPrinted in FranceAcknowledgements: The authors would like to acknowledge Clare O’Hagan forhercontributions to theresearch and analysis.Graphic design: Sonia SavciCover design: Sonia SavciIllustrations: Sonia Savci Artificial Intelligence (AI) holds immense potential tosupport global efforts to reduce environmentalimpactby optimizing energy use, enhancing resource management, and improvingclimate modelingand prediction1. However, the accelerated rise of generative AI, particularly Large Language Models(LLMs), has brought new and urgentresourcechallenges. The exponential growth in computationalpower needed to run these models is placing increasing strain on global energy systems, waterresources, and critical minerals,raising concerns about environmental sustainability, equitable access,and competition over limited resources.Achieving ecological resilience in generative AI is not solely a matter of reducing energy consumption,it is about unlocking broader opportunities, expanding equitable access, and enabling scalable,impactful innovation. A fundamental shift toward AI systems that are “clean by design,” with energyand resource efficiency integrated from the outset, is essential. This requires developing models thatare not only high-performing but also lighter, more efficient, and environmentally sustainable,particularlyas generative AI becomes a foundational layer of our digital infrastructure. Embracingenergy-and resource-efficient AI is key to ensuring that the digital transformation advances in a waythat is both inclusive and ecologically responsible, capable of scaling across diverse global contexts.To turn this vision into reality, addressing generative AI’s sustainability challenges demands sustainedcommitment and collective action. Policymakers, industry leaders, and the scientific community mustprioritize the development of AI systems that are both energy-efficient and accessible,particularly inlow-resource contexts. This report offers three key recommendations to support that shift:(a)mobilize public and private investment, along with strategic partnerships, to drive the developmentand adoption of clean by designAI systems that embed efficiency from the outset; (b) createincentives and standards,such as sustainability labels and green procurement criteria,that encouragetransparency and promote eco-conscious design and usage across the AIecosystem; and (c) enhanceAI literacy to build critical awareness of generative AI’s environmental footprint andfoster moreintentional andconsciousengagement.Through a combination of original experiments and data insights, this report illustrates how practicaltechniques can help translate that vision into action.Methods such as quantization and promptoptimization reduced the energy consumption of large language models by up to 75%withoutcompromisingaccuracy.Moreover, in tasks that arespecializedand repetitive, such as translation orsummarization, replacing a large general-purpose model with smaller, task-specific models led toEXECUTIVE SUMMARY 5 6energy reductions of up to 90%, while maintaining strong performance.These findings offer a tangible,scalable pathway toward a smarter, more accessible, and more resource-efficient AI future.UNESCO’s commitment to advancing the right to information, equitable access to knowledge, andethical digital transformation underpins this report. As generative AI becomes a foundational layer ofdigital infrastructure, it is essential to ensure that itsdevelopment supports sustainability, inclusion,and the public interest. This report provides evidence-based insights into how energy-and resource-efficient AI can help achieve these goals, especially in low-r