Common but Different Futures: AI Inequity and Climate Change
h worse quality public services.”[18] Part of the challenge for low and middle income countries (LMICs) is the absence or unreliable availability of basic infrastructure like electricity and high-speed internet. Similar inequalities mark R&D, patents, startups, funding, skilling and hiring in AI, with the United States and Europe accounting for the lion’s share of investment, academic output, and hiring (see Figures 1a,1b, and 1c).[19] From Top Left: Number of AI Patents; Global distribution of private investments in AI, AI hiring. Source: Global AI Vibrancy Tool, Stanford University. Regional labels are the author’s own.[20] How would the concentration of AI development and capacity—technical and governance—in the Global North affect emissions, and by extension, emission politics and narratives? As Anita Gurumurthy and Nandini Chami of IT for Change write in their 2019 essay:[21] The AI-led global order is entrenched firmly in what activists and scholars have argued is a form of neocolonisation. Today, economic power is a function of how AI technologies are employed in networked systems organised around incessant data processing. As data started flowing on a planetary scale with the advent of the internet, creating and multiplying social and economic connections, predatory capitalism found a new lease of life. The incumbents of the digital revolution, who have shaped global value chains, have the first-mover advantage in AI. This edge is not only in data, digital infrastructure, and capital, but also their ability to set the terms by which other actors engage in governance and ethical debates. (This idea is explored further in the final section.) Elephant in the Dark: Granularity in Emissions In 2020, ICT accounted for between 0.8 and 2.3 gigatons CO2eq in global GHG emissions. Researchers put ICT’s contribution at 1.8 and 2.9 percent of global emissions according to low and mid estimates, and up to 6.3 percent per the “worst-case” estimates.[22] At the same time, AI development and adoption across sectors has skyrocketed, as has compute demand[b] associated with even larger AI models. The compute demand of large AI models, according to a 2018 study by OpenAI, has been doubling every 3.4 months—meaning that since 2012, compute has grown by 300,000 times.[23] Some studies have also attempted to quantify the hypothetical carbon emissions generated by neural network training in different regions, based on server location, type of GPU, and training time.[24] Another study on Natural Language Processing (NLP) models estimated that training a single model generated five times the volume of CO2 emission as a car in its entire lifetime.[25] Table 1: Carbon Footprint of Major NLP Models Estimated carbon costs and cloud compute costs for selected training models. Source: Emma Strubell, Ananya Ganesh and Andrew McCallum, “Energy and Policy Considerations for Deep Learning in NLP”. [26] We can therefore turn to data centre energy use as a partial proxy for AI-related compute demand.[27] A 2015 research paper on ICT-linked electricity consumption estimated energy use of data centres to hit 539 TWh in 2018, and 2967 TWh in 2030, even with improvements in efficiency.[28] A 2017 update to this paper noted that by 2025, data centres could account for 3.2 percent of global carbon emissions.[29] A 2020 study measuring energy use against compute demand, from 2010 to 2018, noted a 6-percent increase in energy use.[30] Global data centre energy use, it further found, accounted for 1 percent of global electricity consumption, which—for comparison—is more than the total electricity consumption of a country like Thailand.[31] Current projections indicate that the APAC data centre market is expected to grow by 12.2 percent between 2020-24, with Southeast Asia alone growing at 12.9 percent. This is followed by Europe, the Middle East, and Africa at 11.1 percent, and North America at 6.4 percent.[32] Figure 2: Regional Growth in Data Centre Markets Regional Data Centre Growth (2020-2024). Source: Cushman and Wakefield (2021)[33] A key issue with many of these data centre studies, however, is that the geographic groupings they employ do not help generate granular insights. Figure 3 demonstrates this using regional data centre statistics from four major cloud service providers—Amazon Web Services, Google Cloud, IBM Cloud, and Microsoft Azure—followed by a breakdown by country in Table 2. Figure 3: Data Centres by Region Regional distribution of data centres of major CSPs. Source: Amazon Web Services Global Infrastructure, Google Data Centers, IBM Cloud, Microsoft Azure.[34] Table 2: Data Centres by Country Select country-level number of data centres of major CSPs. Source: Amazon Web Services Global Infrastructure, Google Data Centers, IBM Cloud, Microsoft Azure.[35] While regional data might be useful in providing high-level insights, such as the paucity of centres in Sub-Saharan Africa and South America, it paints an incomplete picture. For instance, the United States alone accounts for the overwhelming majority of data centres operated by the big four, as well as an average of 39.5 percent of availability zones worldwide.[36] Therefore, even as other regions project double-digit growth in the coming decade, data centre infrastructure is currently unbalanced, and likely to remain so in the near future. In Search of an Equitable Model for Sustainable AI Powerful actors such as governments, regional institutions, and technology companies, have already embarked on a process of building narratives on AI and emissions. This risks recreating the same inequities that have historically marked climate agreements. For instance, technology giants have responded to climate concerns by announcing “net zero” policies and initiatives. Microsoft has pledged to be carbon-negative by 2030 and remove all the carbon the company has emitted since 1975;[37] Alphabet, for its part, has announced sustainability bonds worth USD 5.75 billion that will fund environmentally and socially responsible projects;[38] meanwhile, Facebook is undertaking initiatives toward sustainable supply chains;[39] and Amazon has made a pledge to be net-zero by 2040.[40] To be sure, such pledges are an important signalling tool, indicating that internet giants acknowledge their massive carbon footprint. However, they often rely on the decades-old inequitable carbon offset system which is being criticised for allowing companies to purchase their way out of making any fundamental change in how their operate.[c] There is also little transparency regarding the lifecycle emissions of their operations, including not just the facilities under their direct ownership, but within their broader global supply chains—this makes these “net zero” claims nearly impossible to measure. Another example is the focus on compute efficiency. In January 2021, Google announced the launch of Switch Transformer, a more efficient version of the older, more unwieldy Transformer. The idea with increasing compute efficiency is increasing the number of parameters in a neural network and improving performance, while keeping compute costs constant.[41] Yet, the emphasis on “efficiency” as the silver bullet to offset emissions distracts from the fact that there is still little transparency on the impact of such measures on actual life cycle emissions, leaving independent researchers who may want to verify these claims in the dark. Additionally, efficiency-oriented solutions have the second- and third-order effect of reducing costs and increasing consumption, termed the Jevons Paradox. This is a relationship seen, historically, in ICT-enabled efficiency improvements, where efficiency gains in energy use required for ICT reduced production costs, which led to an increase in the overall consumption of energy.[42] This phenomenon is captured in two relatively new terms—“ethicswashing” and “greenwashing”. Entities seek to mark an ethical checkbox to assuage the concerns of their increasingly climate-conscious shareholders and customers, without undertaking any substantive changes in their global operations.[43] Recommendations Local impact assessment. As earlier sections pointed out, granular data, in terms of geography and energy mix, are needed to drive policy action. Researchers are already proposing models for emissions impact,[44] but require more robust datasets to provide actionable recommendations. Other researchers have also recommended integrating an Environmental, Social, and Governance (ESG) framework in AI governance.[45] Working toward complementary standards for AI emissions governance across geographies. Some geographies are already instating carbon-neutral requirements for data centres. For instance, several CSPs and data centre operators with a presence in Europe—including Google, IBM, AWS, Intel, and Microsoft—have signed a Climate Neutral Data Centre Pact, part of the EU’s roadmap to becoming carbon-neutral by 2050.[46] The Pact sets targets in energy efficiency, transition to clean energy, water conservation, and reuse and repair. These (voluntary) commitments will be monitored by the European Commission. The danger of non-uniform standards is the creation of a new form of “carbon havens”, where global enterprises might move operations to developing countries with comparatively lax regulations on emissions linked to AI and allied technologies. Developing countries should explore the CBDR principle in the context of the climate costs of AI. Developing countries must get ahead of the curve by actively engaging in the process of defining parameters for the climate impact of AI. Small and developing economies are already playing catch-up in AI, contending against powerful incumbents in developed and large economies. While the economic growth imperative of AI is understandably the priority, not engaging in emerging debates in climate and AI risks these narratives and soon, governance processes, being shaped by contexts and terms set by a small group of powerful actors. Annex “Solving” Climate Change: AI for a Sustainable and Inclusive Future 22 February 2021 In February 2021, ORF organised virtual consultations, with a focus on stakeholders from Global South countries. Invited participants included industry representatives, civil society organisations, academia, and relevant government representatives. Discussions centred on AI in the context of the sustainable development goals, specifically SDGs 10 and 13 on reduced inequalities and climate action, respectively. The aim of the consultations was to seek answers to the following questions: How can we forge best practices, mitigate harms, and cooperate in a manner that ensures that the bounties generated by AI will be realised by all, and in a way that leaves a better planet for future generations? Participants put forward four marquee issues and ideas: First, striking the balance between community-centred and state-centred approaches. While the state remains the locus of global governance efforts, a reliance on purely state-centred approaches to sustainable AI will risk marginalising stakeholders whose interests may not be represented at the national level, either because of lack of visibility and resources to make their voices heard or, in some cases, persecution. Second, the lack of interfaces between climate change and AI governance processes. Sustainability needs to become a core principle under ethical AI, and requires the active buy-in of industry, government and multilateral/multistakeholder bodies. Third, making sustainable AI a policy priority for developing countries. The COVID-19 pandemic will further intensify the focus on economic recovery for developing countries, but sustainable recovery—including through sustainable AI—should remain in focus. Finally, the need to acknowledge differential capacities. A common framework for sustainable AI should account for differences in capacity, while balancing the geopolitical framing that characterises global governance of emerging technologies. Participants Abhishek Gupta Founder, Montreal Institute of AI Ethics and Machine Learning Engineer, Microsoft Anirudh Kanisetti Associate Fellow, Takshashila Institution Arthur Vieira Plataforma CIPÓ Attlee Gamundani Young ICTD Fellow, United Nations University Institute in Macau Christopher Cordova Co-founder and Director, AI for Climate Danit Gal Associate Fellow, Leverhulme Centre for the Future of Intelligence Emanuela Girardi Founder, Pop AI & High-Level Expert Group on AI of the Italian Government, Italy Eniola Mafe Lead, 2030 Vision Secretariat, World Economic Forum Gabrielle Alves Junior Researcher, Plataforma CIPÓ Janet Salem Economic Affairs Officer, UN Economic and Social Commission for Asia and the Pacific Marie-Therese Png Ph.D. Candidate, Oxford Internet Institute Olga Cavalli Co-founder/Academic Director, ARGENSIG – SSIG Priya Donti Co-founder, Climate Change AI Serge Stinckwich Head of Research, UN University Institute of Macau Trisha Ray is an Associate Fellow at the Centre for Security, Strategy and Technology at the Observer Research Foundation. Endnotes [a] “The countries or companies that reduce emissions below their cap have something to sell, an unused right to emit, measured in tonnes of CO2 equivalent. Countries and companies that don’t meet their target can buy these one-tonne units to make up the shortfall. This is called emissions trading, or cap and trade.” See: “What are Market and Non-Market Mechanisms?”, UNFCCC [b] ‘Compute demand’ refers to the demand for computational power to carry out computing tasks, such as storage, processing and analytics. [c] Carbon offsetting enables entities to “compensate” for their emissions by funding projects elsewhere that reduce emissions. See: United Nations Carbon Offset Platform: [1] See, for example: David Rolnick, Priya L. Donti, Lynn H. Kaack et al, “Tackling Climate Change with Machine Learning”, arXiv:1906.05433 [cs.CY]. Cedric Villani, “For a Meaningful Artificial Intelligence: Towards a French and European Strategy”, AI for Humanity (2018). “Artificial Intelligence Strategy of the German Federal Government”, The Federal Government of Germany (2020). “Austria AI Strategy Report”, European Commission (2019). [2] “Government AI Readiness Index (2020)”, Oxford Insights. [3] Anil Agarwal and Sunita Narain, “Global Warming in an Unequal World: A Case of Environmental Colonialism”, New Delhi, India: Centre for Science and Environment (1991). [4] ORF Consultations: “Solving” Climate Change: AI for a Sustainable and Inclusive Future”, February 22, 2021 (see Annex). [5] “What is the Kyoto Protocol?”, United Nations Framework Convention on Climate Change, accessed October 1, 2021. “The Cancun Agreements: Outcome of the Work of the Ad Hoc Working Group on Long-term Cooperative Action Under the Convention,” 15 March 2011, Doc. FCCC/CP/2010/7/Add.l. Sikina Jinnah, “Makers, Takers, Shakers, Shapers: Emerging Economies and Normative Engagement in Climate Governance”, Global Governance, Vol. 23, No. 2 (Apr.-June 2017), pp. 285-306. [6] Sikina Jinnah, “Makers, Takers, Shakers, Shapers: Emerging Economies and Normative Engagement in Climate Governance”. [7] See, for instance: Implications of the Kyoto Protocol on Climate Change: Hearing 105-457 Before the Senate Committee on Foreign Relations, 105th Congress (1998). [8] Anirruddh Mohan, “From Rio to Paris: India in global climate politics”, ORF Occasional Paper 130, Observer Research Foundation, December 2017. [9] Anirruddh Mohan, “From Rio to Paris: India in global climate politics”. [10] “Paris Agreement”, UNFCCC (2015). [11] The EU submitted INDCs as a bloc. See: “CAIT Paris Contributions Map”, World Resources Institute. [12] Kate Hampton, Mridula Pandey and Shirish Sinha, “A just path to a decarbonised future”, Observer Research Foundation, November 12, 2021. [13] Lorenzo Fioramonti, How Numbers Rule the World: The Use and Abuse of Statistics in Global Politics (Zed Books, London: 2014): p. 86. [14] Angela Carpenter, and Marcus Wagner, “Environmental justice in the oil refinery industry: A panel analysis across United States counties”, Ecological Economics 159 (2019): 101-109. Heidi Bachram, “Climate fraud and carbon colonialism: the new trade in greenhouse gases”, Capitalism nature socialism 15, no. 4 (2004): 5-20. [15] Alan Chan, Chinasa T. Okolo, Zachary Terner et al, “The Limits of Global Inclusion in AI Development”, arXiv:2102.01265v1 [cs.CY], Feb 2, 2021. [16] “AI Policy and National Strategies”, Artificial Intelligence Index Report 2021, Stanford University. [17] “Government AI Readiness Index (2020)”, Oxford Insights. [18] “Government AI Readiness Index (2020)”, Oxford Insights. [19] Data from “Global AI Vibrancy Tool”, Stanford University. [20] “Global AI Vibrancy Tool”, Stanford University. [21] Anita Gurumurthy and Nandini Chami, “The Wicked Problem of AI Governance”, Artificial Intelligence in India Vol. 2 (2019). [22] Charlotte Freitag, Mike Berners-Lee, Kelly Widdicks et al, “The climate impact of ICT: A review of estimates, trends and regulations “, arXiv (2020). [23] Dario Amodei and Danny Hernandez, “AI and Compute”, OpenAI, May 16, 2018. [24] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt et al, “Quantifying the Carbon Emissions of Machine Learning”, arXiv1910.09700v2 [cs.CY], November 4, 2019. [25] Emma Strubell, Ananya Ganesh and Andrew McCallum, “Energy and Policy Considerations for Deep Learning in NLP”, arXiv:1906.02243v1 [cs.CL], [26] Adapted from Emma Strubell, Ananya Ganesh and Andrew McCallum, “Energy and Policy Considerations for Deep Learning in NLP”, arXiv:1906.02243v1 [cs.CL]. [27] Charlotte Freitag, Mike Berners-Lee, Kelly Widdicks et al, “The climate impact of ICT: A review of estimates, trends and regulations “, arXiv (2020). [28] Anders S.G. Andrae and Tomas Edler, “On Global Electricity Usage of Communication Technology: Trends to 2030” Challenges 6, no. 1: 117-157 (2015). [29] Anders S.G. Andrae, “Total Consumer Power Consumption Forecast”, Nordic Digital Business Summit (October 2017). [30] Eric Masanet, Arman Shehabi, Nuoa Lei, Sarah Smith, and Jonathan Koomey, “Recalibrating global data centre energy-use estimates.”, Science 367, no. 6481 (2020): 984-986. Estimates do vary based on methodology, as there is no set standard for measuring the emissions and energy impact of ICT. See also: Lotfi Belkhir, Ahmed Elmeligi, “Assessing ICT global emissions footprint: Trends to 2040 & Recommendations”, Journal of Cleaner Production, Volume 177(2018): Pages 448-463, [31] “Energy Consumption: International”, U.S. Energy Information Administration, [32] “Data centre Market Global Comparison”, Cushman & Wakefield (2021). [33] “Data centre Market Global Comparison”, Cushman & Wakefield. [34] “Regions and Availability Zones”, Amazon Web Services. “Discover our Data Center Locations”, Google Data Centers. “IBM Cloud global data centers”, IBM Cloud, “Global Infrastructure”, Microsoft Azure, [35] [35] “Regions and Availability Zones”, Amazon Web Services. “Discover our Data Center Locations”, Google Data Centers. “IBM Cloud global data centers”. “Global Infrastructure”, Microsoft Azure. [36] There is also variation not just in per capita electricity consumption, but also in energy mix and usage of data centres amongst locations in the same region, which is masked in grouped data. For instance, cloud servers located in North America emit anywhere from 20g CO2eq/kWh (Quebec, Canada) to 736.6 CO2eq/kWh (Iowa, US). [37] Brad Smith, “Microsoft will be carbon negative by 2030”, Microsoft Blog, January 16, 2020. [38] Ruth Porat, “Alphabet issues sustainability bonds to support environmental and social initiatives”, Google Blog, August 3, 2020. [39] “How Facebook partners with academia to help drive innovation in energy-efficient technology”, Facebook Research, February 5, 2021. [40] “All In: Staying the Course on Our Commitment to Sustainability”, Amazon. [41] William Fedus, Barret Zoph, Noam Shazeer, “Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity”, arXiv:2101.03961 [cs.LG], [42] Miriam Börjesson Rivera, Cecilia Håkansson, Asa Svenfelt and GöranFinnveden, “Including second order effects in environmental assessments of ICT”, Environmental Modelling & Software, 56 (2014): pp.105-115. Ray Galvin, “The ICT/electronics question: Structural change and the rebound effect”, Ecological Economics, 120 (2015): pp. 23-31. [43] Natasha Bernal, “Google, Microsoft and the strange world of corporate greenwashing”, Wired, January 31, 2020. See also: “Ethics Washing”, AI Ethics Living Dictionary. [44] See, for example: Dennis Bouley, “Estimating a Data Center’s Electrical Carbon Footprint”, White Paper 66, Schneider Electric. David Patterson, Joseph Gonzalez, Quoc Le et al, “Carbon Emissions and Large Neural Network Training”, arXiv:2104.10350v3 [cs.LG], April 23, 2021. [45] Abhishek Gupta, Camylle Lanteigne and Sara Kingsley, “SECure: A Social and Environmental Certificate for AI Systems”, arXiv:2006.06217v2 [cs.CY], July 19, 2020. [46] “Signatories”, Climate Neutral Data Centre Pact.
