Revisit the environmental impact of artificial intelligence: the overlooked carbon emission source?

Abstract

The boosting development of artificial intelligence (AI) is contributing to rapid exponential surge of computing power demand, which results in the concerns on the increased energy consumption and carbon emission. To highlight the environmental impact of AI, a quantified analysis on the carbon emission associated with AI systems was conducted in this study, with the hope of offering guidelines for police maker to setup emission limits or studies interested in this issue and beyond. It has been discovered that both industry and academia play pivotal roles in driving AI development forward. The carbon emissions from 79 prominent AI systems released between 2020 and 2024 were quantified. The projected total carbon footprint from the AI systems in the top 20 of carbon emissions could reach up to 102.6 Mt of CO₂ equivalent per year. This could potentially have a substantial impact on the environmental market, exceeding $10 billion annually, especially considering potential carbon penalties in the near future. Hence, it is appealed to take proactive measures to develop quantitative analysis methodologies and establish appropriate standards for measuring carbon emissions associated with AI systems. Emission cap is also crucial to drive the industry to adopt more environmentally friendly practices and technologies, in order to build a more sustainable future for AI.

Graphical abstract

Keywords

Carbon emission / Artificial intelligence (AI) / Training compute / Energy consumption / Economic analysis / Emission caps

References

[1] AchiamJ, Adler S, AgarwalS, AhmadL, AkkayaI, AlemanF L, Almeida D, AltenschmidtJ, AltmanS, Anadkat S J, et al. (2023). Gpt-4 technical report. arXiv: abs/2303.08774.

[2] Argerich M F, Patiño-Martínez M. (2024). Measuring and improving the energy efficiency of large language models inference. IEEE Access: Practical Innovations, Open Solutions, 12: 80194–80207

[3] Bozeman J F (2024) Bolstering integrity in environmental data science and machine learning requires understanding socioecological inequity. Frontiers of Environmental Science & Engineering, 18(5): 65

[4] BrownT B, Mann B, RyderN, SubbiahM, KaplanJ, DhariwalP, Neelakantan A, ShyamP, SastryG, AskellA, et al. (2020). Language Models are Few-Shot Learners. ArXiv: abs/2005.14165

[5] Chen W, Zhang Q, Hu L, Geng Y, Liu C. (2023). Understanding the greenhouse gas emissions from China’s wastewater treatment plants: based on life cycle assessment coupled with statistical data. Ecotoxicology and Environmental Safety, 259: 115007

[6] ChienA A, Lin L, NguyenH, RaoV, SharmaT, WijayawardanaR (2023). Reducing the carbon impact of generative AI inference (today and in 2035). In: Proceedings of the 2nd Workshop on Sustainable Computer Systems, Association for Computing Machinery, New York, USA, Article 11, 1–7

[7] de Vries A. (2023). The growing energy footprint of artificial intelligence. Joule, 7(10): 2191–2194

[8] Energy Institute (2024). Statistical Review of World Energy. Edinburgh. Heriot-Watt University Available online at the website of energyinst.org (accessed Augest 02, 2024)

[9] EpochAI (2024). Compute Trends Across Three Eras of Machine Learning. San Francisco. Rethink Priorities. Available online at the website of epochai.org/blog/compute-trends (accessed April 24, 2024)

[10] HannahR, Pable R, MaxR (2020). Greenhouse gas emissions. Our World in Data. Available online at the website of ourworldindata.org/greenhouse-gas-emissions (accessed Augest 02, 2024)

[11] HuangHMa RRenH (2024). Scientific and technological innovations of wastewater treatment in China. Frontiers of Environmental Science & Engineering 18(6): 72

[12] LacosteA, Luccioni A S, SchmidtV, DandresT (2019). Quantifying the carbon emissions of machine learning. ArXiv: abs/1910.09700.

[13] LuersA, Koomey J, MasanetE, GaffneyO, Creutzig F, Lavista FerresJ, HorvitzE (2024). Will AI accelerate or delay the race to net-zero emissions? Nature, 628(8009): 718–720

[14] OWID (2024). Per capita CO₂ emissions. Oxfordshire. Global Change Data Lab. Available online at the website of ourworldindata.org/grapher/co-emissions-per-capita (accessed April 25, 2024)

[15] Rillig M C, Ågerstrand M, Bi M, Gould K A, Sauerland U. (2023). Risks and benefits of large language models for the environment. Environmental Science & Technology, 57(9): 3464–3466

[16] ScellsH, Zhuang S, ZucconG (2022). Reduce, reuse, recycle. In: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Association for Computing Machinery, New York, NY, USA, 2825–2837

[17] SevillaJ, Heim L, HoA C, BesirogluT, Hobbhahn M, VillalobosP (2022). Compute Trends Across Three Eras of Machine Learning. International Joint Conference on Neural Networks (IJCNN), 1–8

[18] Similarweb(2024). Website Analysis. New York. Similarweb Ltd. Available online at the website of pro.similarweb.com/#/digitalsuite (accessed August 27, 2024).

[19] WB (20232023. State and Trends of Carbon Pricing 2023. Washington, D.C. World Bank. Available online at the website of hdl.handle.net/10986/39796 (accessed April 25, 2024)

[20] Zhu J J, Jiang J, Yang M, Ren Z J. (2023). ChatGPT and environmental research. Environmental Science & Technology, 57(46): 17667–17670

RIGHTS & PERMISSIONS

Higher Education Press 2024