Generative AI exists within a complex web of human activities that collectively impact our climate.
While concerns about AI's rapidly expanding carbon and water footprint are valid, they demand proper contextualization alongside other common activities (air conditioners, electric cars, data centers).
The carbon impact of daily ChatGPT usage (approximately 4-40 kg CO₂/year for moderate-heavy use) appears relatively modest when compared to activities like transatlantic flights (approx. 3.2 tonnes CO₂ per roundtrip), car ownership (~ 4.6 tonnes CO₂/year), or crypto (in 2021, Bitcoin produced ~ 65.4 million tonnes of CO₂e; in 2025, it increased in impact to 91.4 MtCO₂e, according to Cambridge Blockchain Network Sustainability Index (CBNSI)), and the military (the US military in 2017 exuded ~193,414 metric tons of CO₂e per day).
The social utility of war (in an age of nuclear weapons) is zero.
Linear scale comparison of carbon footprints [estimates] across human activities
| Activity | Emission Estimate | Source (APA + link) | Data Location · Math Synopsis |
|---|---|---|---|
| Military (U.S. per day) | ~547.9 kg CO₂e soldier⁻¹ day⁻¹; 25 – 40 t CO₂e soldier⁻¹ yr⁻¹ |
Crawford, N. C. (2019). Pentagon Fuel Use, Climate Change, and the Costs of War. Costs of War Project, Brown Univ. Link |
Math. 200 t yr⁻¹ lifecycle ÷ 365 ≈ 547.9 kg day⁻¹; 1.2 Gt (2001-17) ≈ 193 kt day⁻¹. ~25 t CO₂e yr⁻¹ is a conservative per-soldier estimate focused on direct operational output. ~51 t yr⁻¹ (from total divided by active-duty count) better reflects broad operational plus some indirect emissions. True full-spectrum impact (Scopes 1–3, including economic and ecological destabilization) would likely exceed both. ~547.9 kg CO₂e per soldier per day is based on a comprehensive estimate of 200 t CO₂e yr⁻¹ … “The U.S. Department of Defense is the largest institutional consumer of fossil fuels … emitted 1.2 billion t CO₂e.” Total U.S. military: ~193 400 t CO₂e per day. |
| Bitcoin ASIC miner | ~9 t CO₂e yr⁻¹ | de Vries, A. (2018). Bitcoin’s Growing Energy Problem. Joule, 2(7), 801-805. Link |
Math. 22–23 Mt CO₂e yr⁻¹ ÷ ≈ 2.5 M ASICs ≈ 9 t device⁻¹ yr⁻¹. Estimate based on network-wide CO₂e emissions and approximate per-device allocation. Stoll (2019) estimated Bitcoin’s annual electricity use at ~45.8 TWh, producing 22.0–22.9 Mt CO₂. The ~9 t yr⁻¹ per ASIC miner reflects a heuristic division by the estimated global fleet of continuously operating mining devices. |
| Gasoline car (12 000 mi yr⁻¹) |
~4.6 t CO₂e yr⁻¹ | United States Environmental Protection Agency. (2024). Greenhouse Gas Emissions from a Typical Passenger Vehicle. Link |
Math. EPA average 4.6 t yr⁻¹ for 11 500 mi. Stated average: 4.6 t CO₂e yr⁻¹ for 11 500 miles. |
| Transatlantic flight (RT, economy) | ~2.9 t CO₂e trip⁻¹ (RF × 2) |
atmosfair. (n.d.). Flight CO₂ Calculator & Climate Impact Methodology. Calculator · Method |
Math. CO₂ only ≈ 1.45 t (LHR-JFK RT); atmosfair factor 2 ⇒ 2.9 t. Calculator result for round-trip London–New York, economy class. atmosfair: “the effects of CO₂, contrails, ozone, etc. drive global warming two to five times more than CO₂ alone … atmosfair relates CO₂ to the other effects using an average factor of 3 …” (German Federal Environmental Agency, 2008). |
| Generative AI (1 000 images day⁻¹) | ≈ 40 – 1 100 kg CO₂e yr⁻¹ |
Wen, S., Liu, B., Shao, Y., & Wood, A. (2023). BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion. arXiv 2311.16863.
Link Luccioni, A. S., Jernite, Y., & Strubell, E. (2024). Power Hungry Processing: Watts Driving the Cost of AI Deployment? Proc. FAccT ’24, 85-99. PDF Stability AI. (2024). Introducing Stable Diffusion 3.5. News |
Math. 0.1 g image⁻¹ ⇒ 40 kg yr⁻¹; 1.6 g ⇒ 584 kg; extrapolated 3 g ⇒ 1.1 t. Grid 400-600 g kWh⁻¹; training + <5 %. Measured inference intensities range from 0.1 g CO₂e image⁻¹ (segmind/tiny-sd, 500 M params) to 1.6 g (SD-XL, 2.6 B). SD 3.5 Large (8.1 B) lacks peer-reviewed data; scaling ∝ N1.4 places it near 3 g image⁻¹. At 1 000 images day⁻¹ × 365 days this yields ≈ 40 kg (tiny-sd) to ≈ 1.1 t (SD 3.5) CO₂e yr⁻¹. Figures are inference-only; training amortisation < 5 %. |
| Video streaming | ~55 g CO₂e h⁻¹; 40 kg yr⁻¹ (2 h day⁻¹) |
Carbon Trust. (2021). Carbon Impact of Video Streaming. Link |
Math. 55 g h⁻¹ × 2 h day⁻¹ × 365 = 40 kg yr⁻¹. Page 8: “The European average footprint … approximately 55 g CO₂e per hour of video streaming for the conventional allocation approach.” |
| ChatGPT | ~2.2 g CO₂e query⁻¹; ~40 kg yr⁻¹ (50 queries day⁻¹) |
Tomlinson, B., Black, R. W., Patterson, D. J., & Torrance, A. W. (2024). The Carbon Emissions of Writing and Illustrating Are Lower for AI than for Humans. Scientific Reports, 14, 3732. DOI |
Math. 552 t training ÷ 300 M queries = 1.84 g; + 0.38 g inference = 2.2 g query⁻¹. This is a severe approximation. Many other estimates are far lower. Estimate is mid-to-high range; incorporates monthly model re-training into each inference. Results: “552 metric t ÷ 300 M queries = 1.84 g … combined impact ≈ 2.2 g CO₂e per query.” |
Note: All these are estimates. These do not incude embodied emissions, the cradle-to-grave pathways. Little is quantitatively possible to derive exactly from such sprawling complex planetary systems. But, pause to consider that an AI user generating 1,000 images per day for an entire year will cumulatively generate less than the climate impact of feeding a large pet dog for a year, or taking a single transatlantic flight.
What does the data say?
"In 2022, global data center electricity consumption was estimated to be between 240 and 340 terawatt-hours (TWh), accounting for approximately 1-1.3% of global final electricity demand. This excludes energy used for cryptocurrency mining, which was estimated to be around 110 TWh in 2022." -- IAE, Global Conference on Energy & AI December 2024
The U.S. Department of Defense emitted ≈ 1.2 billion metric tons CO₂-equivalent during FY 2001-2017 (direct operational and installation sources only). In FY 2006 it purchased about 30 TWh of electricity. Global military greenhouse-gas emissions are still larger once fuel and other energy-intensive activities across all armed forces are counted. — Pentagon Fuel Use, Climate Change, and the Costs of War (Watson Institute)
Our World in Data reports: "Non-CO2 climate impacts mean aviation accounts for around 4% of global warming to date. While aviation accounts for around 2.5% of global CO2 emissions, its overall contribution to climate change is higher. Along with emitting CO2 from burning fuel, planes also affect the concentration of other atmospheric gases and pollutants. They generate a short-term increase but a long-term decrease in ozone and methane and increased emissions of water vapor, soot, sulfur aerosols, and water contrails."
Note:“Aviation” slice shown for climate context although its energy is liquid fuel, not electricity. "AI-accelerated servers" are not necessarily dedicated exclusively to AI. “All other sectors” = industry, buildings, non-AI ICT, transport, public services, agriculture, etc.
Sources: IEA Data Centres & Data Transmission Networks 2023; IEA Energy & AI, Apr 2025; Digiconomist Bitcoin index (110 TWh ≈ 0.4 %); Brown University Costs of War (DoD energy ⇢ global military ≈ 0.4 % power); Our World in Data aviation dataset (2.5 % CO₂, ~4 % warming).
"Currently the global average consumption of web surfing, social media, video and music streaming, and video conferencing could account for approximately 40% of the per capita carbon budget consistent with limiting global warming to 1.5 °C" -- Source: Istrate, R., Tulus, V., Grass, R. N., Prévôt, A. S. H., & Wick, P. (2024). The environmental sustainability of digital content consumption. Nature Communications, 15, 3724. https://doi.org/10.1038/s41467-024-47621-w.
As SemiAnalysis states: "The carbon emissions from these training runs are significant, with one GPT-3 training run generating 588.9 metric tons of CO2e, equivalent to the annual emissions of 128 passenger vehicles. Complaining about GPT-3 training emissions is like recycling plastic water bottles but then taking a flight every few months. Literally irrelevant virtue signaling." Source: Patel, D., Nishball, D., & Ontiveros, J. E. (2024, March 13). AI Datacenter Energy Dilemma – Race for AI Datacenter Space. SemiAnalysis. https://semianalysis.com/2024/03/13/ai-datacenter-energy-dilemma-race/
A fully booked Boeing 777 with 300+ passengers produces more CO₂e in a single trantlantic flight than one training run of a GPT-3 size model: the run equals ≈ 60 – 70 % of one full flight.
As of 2022, data centres (excluding cryptocurrency mines) used an estimated 240 – 340 TWh of electricity—about 1 – 1.3 % of global demand (IEA). That power underpins the internet’s core services: web hosting, social media, streaming video, on-line gaming, Wikipedia and more. It is growing more energy-efficient as it expands.
Proof-of-work cryptocurrency (Bitcoin, etc.) mining —whose net social value I regard as marginal and whose main effect is to amplify speculative greed— burned a further ~110 TWh in 2022, roughly 0.4 % of global electricity and about one-third of non-crypto data-centre use (IEA).
If we could fully price the carbon, electricity, concrete and post-conflict reconstruction bound up in global military activity, its footprint would likely sit in a comparable range (1/3 of global data centre impact). The social utility of war (in an age of nuclear weapons) is zero. (Brown University Costs of War; Wikipedia).
Commercial aviation emits ~2.5 % of anthropogenic CO₂ yet, once contrails and other non-CO₂ effects are counted, contributes about 4 % to global warming (Our World in Data).
By 2025 AI-accelerated servers are projected to draw roughly one-fifth of all data-centre electricity—about 0.4 % of global power—overtaking cryptocurrency mining yet still trailing conventional cloud and streaming workloads (Source: IEA April 2025 report on Energy & AI which is "the first comprehensive global analysis examining all aspects of the links between energy and AI – from pathways to securely and sustainably meeting energy demand for AI, to how AI itself could transform the production, consumption and transport of energy around the world. The analysis explores the implications of the rise of AI on energy security, investment, emissions and more – providing a strong factual basis for those thinking through the challenges and opportunities ahead ... How much electricity demand comes from AI specifically is a challenging question to answer. AI is only one of the workloads that run on data centres, and as AI becomes increasingly pervasive, a clear distinction between AI-related and non-AI-related workloads becomes more challenging.")
Note how in this chart of energy usage, dedicated AI (colored blue) is still a fraction of crypto. 
AI is still as of 2025 only a small slice of the catastrophe. It will grow exponentionally. But consider what it contributes: knowledge, medicine, proteins, material discovery, optimizations, code, etc etc etc. It is a profoundly useful technology.
Each emissions category represents a distinct metabolic process within our planetary system, characterized by specific patterns of molecular exchange and transformation. The following analysis examines these processes in detail.
≈ 25,000 - 50,000 kg CO₂e per soldier annually
Per soldier in the United States military establishment
This measurement encompasses the carbon metabolism associated with contemporary military operations, including direct emissions from vehicle and aircraft operation, facility maintenance, and field operations, as well as supply chain emissions from equipment manufacturing, transport logistics, and administrative activities.
The exceptional carbon intensity of military activity emerges from multiple factors: the operation of specialized high-performance vehicles and aircraft optimized for capabilities rather than efficiency; the maintenance of redundant systems for operational security; the privileging of logistical certainty over resource optimization; and the inherently energy-intensive nature of contemporary weapons systems.
Note: This value represents an estimate based on aggregate military fuel consumption and operational data, divided by total personnel count. Actual emissions vary significantly across different military branches and deployment scenarios.
≈ 9,000 kg CO₂e (Mining) / ≈ 3,600 kg CO₂e (Transactions) annually
Mining: Small-scale operation with one ASIC machine / Transactions: 12 monthly Bitcoin transfers
These measurements reflect two distinct modalities of cryptocurrency engagement. Transaction emissions derive from the distributed consensus mechanisms that validate and secure blockchain operations, where hundreds of thousands of computers simultaneously perform cryptographic operations to reach agreement on ledger state. The mining figure represents the continuous operation of specialized hardware designed exclusively for proof-of-work computation.
The carbon intensity of these processes varies substantially across global regions, with some mining operations strategically positioned near renewable energy sources (particularly hydroelectric facilities with excess capacity) while others exploit low-cost fossil fuel energy.
Note: Cryptocurrency emissions are highly dynamic and vary with network hashrate, energy sources, and market conditions. These figures represent estimates from peer-reviewed research and may not reflect the most current efficiency improvements or network changes.
Cambridge Bitcoin Electricity Consumption Index (2024). & Krause, M. J., & Tolaymat, T. (2018). "Quantification of energy and carbon costs for mining cryptocurrencies." Nature Sustainability, 1(11), 711-718.
≈ 4,600 kg CO₂e (Gasoline) / ≈ 2,800 kg CO₂e (Hybrid) annually
Based on 12,000 miles of annual driving
These figures represent the operational emissions from fuel combustion and do not incorporate the embodied carbon within vehicle manufacturing processes. The substantial differential (approximately 39% reduction with hybrid technology) reflects the metabolic efficiency gained through energy recuperation systems. The gasoline vehicle's carbon signature derives primarily from the thermodynamic inefficiency of internal combustion, where approximately 80% of fuel energy dissipates as waste heat rather than kinetic movement.
Both figures assume standard driving patterns across diverse topographies and traffic conditions, with actual emissions fluctuating based on driving behavior, maintenance practices, and vehicle-specific variables.
EPA. (2023). "Greenhouse Gas Emissions from Typical Passenger Vehicles." Environmental Protection Agency. & US Department of Energy. (2023). "Emissions from Hybrid and Plug-In Electric Vehicles." Alternative Fuels Data Center.
≈ 3,700 kg CO₂e annually
Incremental household emissions attributable to one child in economically developed regions
This measurement represents the additional carbon burden associated with raising a child within contemporary socioeconomic systems, encompassing increased household energy consumption, transportation requirements, nutritional provisions, material goods, and institutional services (education, healthcare).
This figure is deliberately conservative, derived from methodologies that attribute only direct incremental emissions rather than calculating theoretical lifetime impacts or proportional responsibility for systemic emissions. The metric should be understood not as a critique of reproduction but as a recognition of the metabolic intensification that accompanies contemporary child-rearing practices within consumer-oriented societies.
Note: This value represents incremental household emissions only and is based on economic modeling. It does not include proportional attribution of broader societal emissions or lifetime consumption patterns. Substantial variation exists based on socioeconomic position, geographical context, and parenting approaches.
≈ 2,893 kg CO₂e (1 passenger)
One round-trip flight between London and New York (economy class)
This measurement encompasses the direct carbon emissions from jet fuel combustion attributed on a per-passenger basis. The figure represents a standard economy-class allocation for a round-trip journey of approximately 6,900 kilometers (3,450 km each way). Aviation emissions carry additional climate significance due to their release at high altitude, where complex atmospheric chemistry and radiative forcing mechanisms potentially amplify their warming effect beyond the simple carbon dioxide equivalent.
Significant variation exists based on aircraft type, passenger load factor, flight routing, and cabin class (with premium seating allocations carrying proportionally higher emissions).
Atmosfair. (2024). "CO₂ Calculator for Air Travel Emissions." Atmosfair gGmbH.
≈ 2,620 kg CO₂e (Meat-Rich) / ≈ 1,055 kg CO₂e (Vegan) annually
Based on typical consumption patterns in economically developed regions
These measurements encompass the complete lifecycle carbon footprint of food systems: agricultural production, processing, transport, retail, and eventual preparation. The substantial differential (approximately 60% reduction with plant-based nutrition) emerges primarily from the thermodynamic inefficiency of trophic conversion, wherein animal bodies metabolize plant energy with considerable loss. Ruminant animals (cattle, sheep) contribute further to this differential through enteric fermentation producing methane, a potent greenhouse gas with 28-36 times the warming potential of CO₂ over a 100-year period.
These figures represent statistical averages across diverse consumption patterns, with actual emissions varying based on specific food choices, sourcing practices, and waste behaviors.
Scarborough, P., et al. (2014). "Dietary greenhouse gas emissions of meat-eaters, fish-eaters, vegetarians and vegans in the UK." Climatic Change, 125(2), 179-192. & Poore, J., & Nemecek, T. (2018). "Reducing food's environmental impacts through producers and consumers." Science, 360(6392), 987-992.
≈ 1,750 kg CO₂e (Large Dog) / ≈ 350 kg CO₂e (Cat) annually
Based on standard pet ownership practices in developed economies
These measurements encompass the carbon footprint of companion animal nutrition, healthcare, waste management, and associated product consumption. The primary contributor to this carbon signature is animal nutrition, particularly the meat-based protein content of commercial pet foods. The substantial differential between canine and feline emissions emerges from differences in body mass, metabolic requirements, and typical nutritional regimes.
The figures presented represent median values for typical pet ownership practices, with considerable variation based on animal size, diet quality, healthcare intensity, and owner consumption patterns. Recent innovations in insect-based and plant-protein pet foods offer potential pathways for emissions reduction without compromising animal health.
≈ 29 – 41 kg CO₂e (Light) / ≈ 0.18 – 1.1 t CO₂e (Heavy) per year
Light: 50 text queries + 1 image day⁻¹ | Heavy: 1 000 images day⁻¹
Intensity factors used
• Text: ChatGPT-equivalent ≈ 2.2 g CO₂e query⁻¹ (training amortised) and BLOOM ≈ 1.6 g query⁻¹.
• Images: 0.1 g image⁻¹ (segmind/tiny-sd, 500 M params); 1.6 g image⁻¹ (Stable Diffusion XL, 2.6 B); extrapolated ≈ 3 g image⁻¹ (Stable Diffusion 3.5 Large, 8.1 B).
Assumed grid carbon intensity 400–600 g kWh⁻¹; training adds <5 % and hardware embodied carbon <1 %, both omitted.
Math check
Light low – high (50 × 1.6 + 0.1) g to (50 × 2.2 + 3) g day⁻¹ ≈ 80 – 113 g → 29 – 41 kg yr⁻¹.
Heavy low – high (0.5 – 3) g image⁻¹ × 1 000 image day⁻¹ → 0.18 – 1.1 t yr⁻¹.
Coal-heavy grids push the upper bound; renewables push toward the lower bound.
Context
These measurements encompass the computational energy required for neural network inference operations across different usage intensities. The figures derive from specific architectural implementations—BLOOM text model (0.3g CO₂e per query) and Stable Diffusion image generation (2.9g CO₂e per image)—representing current technological efficiency frontiers rather than industry averages.
The carbon signature of AI computation emerges from the intensive matrix operations necessary for pattern recognition and generation, with each inferential process requiring millions to billions of mathematical operations. These operations materialize as electrical current flowing through semiconductor pathways, with attendant thermodynamic inefficiencies manifesting as waste heat requiring further energy expenditure for cooling systems.
Sources
Luccioni, A. S., Jernite, Y., & Strubell, E. (2024). Power Hungry Processing: Watts Driving the Cost of AI Deployment? Proceedings of the 2024 ACM Conference on Fairness, Accountability & Transparency (FAccT ’24), 85–99. PDF
Patterson, D., Gonzalez, J., Le, Q., et al. (2021). Carbon Emissions and Large Neural Network Training. arXiv:2104.10350. Link
Kim, B-K., Song, H-K., Castells, T., & Choi, S. (2023). BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion. arXiv:2305.15798. Link
Stability AI (2024). Introducing Stable Diffusion 3.5. Company news release describing the 8.1 B-parameter SD-3.5 Large and Turbo variants. Link
≈ 40 kg CO₂e annually
Based on 2 hours of daily streaming throughout the year
This measurement encompasses the totality of computational infrastructure necessary to deliver digital video content: data center energy consumption, network transmission requirements, and local device operation. Regional variation is significant, with carbon intensity determined by the underlying energy grid. The European average (55g CO₂e/hour) represents a median value, with substantially higher emissions in coal-dependent regions and lower figures in areas powered by renewable energy sources.
The phenomenological experience of streaming—seemingly immaterial and ephemeral—obscures a complex material infrastructure of server farms, fiber-optic networks, and semiconductor technologies operating continuously to maintain this flow of digital information.
Carbon Trust. (2021). "Carbon impact of video streaming." Carbon Trust.
The visualization and categorical analysis above reveal not merely quantitative differences but qualitative dimensions of our collective participation in a planetary metabolism. Carbon molecules, circulating through these diverse systems of human activity, trace the outlines of technological, economic, and social priorities. Their movements—from military operations to digital communications—constitute a form of planetary writing that transcends territorial demarcations.
A hierarchy of impact emerges that invites reconsideration of the molecular consequences of collective choices. The technological infrastructures that undergird military capabilities (aggression) and speculative capital (greed) reveal themselves as particularly intensive sites of molecular transformation, while newer computational systems—despite their cultural prominence—remain comparatively modest in their carbon circulation.
This quantative exploration arises in conjunction with the Everyone at Home Everywhere project, a one-year durational non-performance exploring planetary belonging through intentional linguistic and behavioral shifts: not using nation-state names, not watching films with weapons.
For further exploration of AI's relationship with ecological consciousness, see the companion project Whole-Use-AI, which examines the ethical and material dimensions of computational systems in creative practice.
Carbon molecules, like all of Earth's material substrates, recognize no territorial demarcations. Their movements and accumulations constitute a planetary metabolism that transcends artificial constructs of separation.
Belcher, Oliver, Patrick Bigger, Ben Neimark, and Cara Kennelly. 2019. “Hidden Carbon Costs of the ‘Everywhere War’: Logistics, Geopolitical Ecology, and the Carbon Boot-Print of the US Military.” Transactions of the Institute of British Geographers 44 (1): 65–80. https://doi.org/10.1111/tran.12319.
Cambridge Centre for Alternative Finance. 2024. “Cambridge Bitcoin Electricity Consumption Index.” University of Cambridge. Accessed March 9, 2025. https://ccaf.io/cbnsi/cbeci/ghg.
Clark, Michael A., Marco Springmann, Jason Hill, and David Tilman. 2019. “Multiple Health and Environmental Impacts of Foods.” Proceedings of the National Academy of Sciences 116 (46): 23357–62. https://doi.org/10.1073/pnas.1906908116.
Crawford, Neta. 2019. “Pentagon Fuel Use, Climate Change, and the Costs of War.” Watson Institute for International and Public Affairs, Brown University. https://watson.brown.edu/costsofwar/papers/ClimateChangeandCostofWar.
De Vries, Alex. 2018. “Bitcoin’s Growing Energy Problem.” Joule 2 (5): 801–5. https://www.sciencedirect.com/science/article/pii/S2542435118301776.
EPA (Environmental Protection Agency). 2024. “Greenhouse Gas Emissions from a Typical Passenger Vehicle.” Accessed March 9, 2025. https://19january2017snapshot.epa.gov/greenvehicles/greenhouse-gas-emissions-typical-passenger-vehicle-0_.html.
International Energy Agency. 2020. "Data Centres and Data Transmission Networks." Accessed March 9, 2025. https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks.
Krause, Max J., and Thabet Tolaymat. 2018. “Quantification of Energy and Carbon Costs for Mining Cryptocurrencies.” Nature Sustainability 1 (11): 711–18. https://doi.org/10.1038/s41893-018-0152-7.
Okin, Gregory S. 2017. “Environmental Impacts of Food Consumption by Dogs and Cats.” PLoS ONE 12 (8): e0181301. https://doi.org/10.1371/journal.pone.0181301.
Poore, Joseph, and Thomas Nemecek. 2018. “Reducing Food’s Environmental Impacts through Producers and Consumers.” Science 360 (6392): 987–92. https://doi.org/10.1126/science.aaq0216.
Toews, Rob. 2019. “Deep Learning's Carbon Emissions Problem.” Forbes. Accessed March 9, 2025. https://www.forbes.com/sites/robtoews/2019/06/06/deep-learnings-carbon-emissions-problem/.
Heikkilä, Melissa. 2023. “AI's Carbon Footprint Is Bigger than You Think.” MIT Technology Review. Accessed March 9, 2025. https://www.technologyreview.com/2023/05/23/1073117/ai-carbon-footprint/.
Erdenesanaa, Delger. 2023. “A.I. Could Soon Need as Much Electricity as an Entire Country.” The New York Times. Accessed March 9, 2025. https://www.nytimes.com/2023/10/10/climate/ai-could-soon-need-as-much-electricity-as-an-entire-country.html.
Desislavov, Radosvet, Fernando Martínez-Plumed, and José Hernández-Orallo. 2023. “Trends in AI Inference Energy Consumption: Beyond the Performance-vs-Parameter Laws of Deep Learning.” Sustainable Computing: Informatics and Systems 1 (April): 100678. https://ui.adsabs.harvard.edu/abs/2023SCIS...3800857D/abstract.
Sundberg, Niklas. 2023. “Tackling AI's Climate Change Problem.” IEEE Spectrum. Accessed March 9, 2025. https://spectrum.ieee.org/ai-carbon-footprint.
Saenko, Kate. 2023. “A Computer Scientist Breaks Down Generative AI's Hefty Carbon Footprint.” The Conversation. Accessed March 9, 2025. https://www.scientificamerican.com/article/a-computer-scientist-breaks-down-generative-ais-hefty-carbon-footprint/
Wynes, Seth, and Kimberly A. Nicholas. 2017. “The Climate Mitigation Gap: Education and Government Recommendations Miss the Most Effective Individual Actions.” Environmental Research Letters 12 (7): 074024. https://doi.org/10.1088/1748-9326/aa7541.
AI Climate links on Zotero or in report format.