European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
The Hidden Environmental Cost of
Generative AI: When Viral Trends Meet
Computational Reality
Abhinay Sama
Indian Institute of Technology Madras, India
doi: https://doi.org/10.37745/ejcsit.2013/vol13n334353 Published June 04, 2025
Citation: Sama A. (2025) The Hidden Environmental Cost of Generative AI: When Viral Trends Meet Computational
Reality, European Journal of Computer Science and Information Technology,13(33),43-53
Abstract: Artificial intelligence has revolutionized human-technology interactions through conversational
interfaces, voice assistants, and image generation capabilities. While these innovations offer remarkable
convenience and efficiency gains across industries, they conceal significant environmental implications
that remain largely invisible to end-users. The disconnect between simple actions like clicking a button and
the substantial computational resources required to fulfill these requests creates an abstraction gap that
obscures environmental consequences. This gap widens as economic incentives drive technological
advancement without proportional consideration for sustainability. The viral #Ghibli trend exemplifies how
social media can rapidly amplify resource-intensive AI features, creating substantial energy demand spikes
before mitigation measures can be implemented. Addressing these challenges requires multifaceted
approaches, including transparent environmental impact indicators, carbon-aware rate limiting,
architectural innovations for efficiency, and enhanced user education. The relationship between
technological progress and environmental responsibility demands greater intentionality in the design and
implementation of AI systems. Making the invisible environmental footprint visible represents a critical
step towards ensuring generative AI enhances human experience without undermining ecological systems,
particularly as these technologies become increasingly embedded in daily digital interactions.
Keywords: Generative AI, environmental impact, computational resources, energy consumption,
sustainability
INTRODUCTION
In the rapidly evolving landscape of artificial intelligence, we've witnessed a significant shift in how people
interact with technology. Traditional search is giving way to conversational AI, with market projections
indicating substantial growth in the coming years [1]. These conversational systems now handle most
43
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
customer interactions without human involvement, dramatically transforming the customer service
landscape across industries. The efficiency gains are substantial, with businesses reporting considerable
cost reductions in customer service operations after implementing conversational AI solutions.
Voice assistants are becoming increasingly sophisticated, with AI-powered voice technology now emerging
as a critical tool for environmental sustainability. Research indicates that voice AI systems can significantly
reduce carbon emissions compared to traditional call centers, primarily by eliminating the need for physical
infrastructure and commuting staff [2]. These voice technologies are processing numerous voice commands
monthly across devices while helping organizations reduce paper usage through digitized documentation
processes. Similarly, image generation capabilities have transformed creative processes across multiple
industries. This technological revolution has fundamentally altered user interaction patterns, with studies
showing that conversational interfaces can reduce user effort compared to traditional graphical interfaces
while increasing customer satisfaction scores.
While these advancements offer unprecedented convenience and capabilities, they come with a hidden cost
that's rarely discussed in mainstream conversations: their environmental impact. The computational
resources required to power these AI systems have grown exponentially, raising urgent questions about
long-term sustainability. Despite voice AI's relative efficiency compared to traditional call centers, the
broader AI sector faces significant environmental challenges, particularly with computationally intensive
applications like image generation and large language models. The paradox emerges where certain AI
implementations can reduce environmental impact in specific domains, while the overall AI infrastructure
continues to demand increasing energy resources. The ecological implications extend beyond mere energy
consumption. Voice AI solutions have demonstrated potential to reduce business travel, lowering associated
carbon emissions while maintaining effective communication channels [2]. Meanwhile, conversational AI
platforms are transforming how customers engage with sustainability-focused products, with consumers
more likely to purchase from companies emphasizing environmental responsibility in their messaging [1].
While training represents a significant one-time environmental cost, the cumulative impact of inference
(serving AI responses to users) poses an equally substantial long-term environmental challenge. As
traditional search increasingly gives way to AI-powered interactions, the energy required to host and run
these models for billions of daily queries creates an ongoing environmental burden that scales with
adoption. A single image generation request may consume 10-15 times more energy than a traditional
search query, while complex reasoning tasks in LLMs can require substantial computational resources for
each user interaction. This inference cost, though less visible than training, accumulates significantly as
these systems become integrated into daily digital experiences.
As we navigate this complex relationship between technological advancement and environmental
responsibility, the need for greater awareness and intentional design becomes increasingly apparent. The
invisible nature of AI's environmental footprint makes this challenge particularly complex, requiring
innovative approaches to measurement, transparency, and sustainable implementation.
44
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
Table 1: Conversational AI Impact [1, 2]
Impact Area Voice AI Systems
Carbon Emissions (% reduction) 95%
Paper Usage Reduction 70%
Customer Interactions without Human Involvement 85%
Customer Service Cost Reduction 30%
Business Travel Reduction 35%
The Invisible Resource Consumption of AI
Large Language Models (LLMs) like GPT-4, Claude, and others require enormous computational
resources, both during training and deployment. OpenAI's GPT-3, a predecessor to the current state-of-the-
art models, was estimated to have consumed approximately 1,287 MWh of electricity during training,
equivalent to the annual energy consumption of 120 US homes. Current models are significantly larger and
more resource-intensive.
The environmental impact of AI extends far beyond the initial training phase. Research on natural language
processing models shows that the carbon footprint associated with developing and tuning these models can
exceed 626,000 pounds of CO₂ emissions, equivalent to nearly five times the lifetime emissions of an
average American car (including manufacturing) [3]. This analysis, which examined models with neural
architecture search, revealed a computational cost of approximately 284 hours on 196 GPUs, with the model
emitting as much carbon as 125 round-trip flights between New York and Beijing.
These figures become particularly concerning when considering the current trajectory of AI research, which
emphasizes ever-larger models to achieve benchmark improvements. Empirical data indicates that the
computational resources needed to produce state-of-the-art AI systems have been increasing exponentially,
doubling approximately every few months since 2012. This rate of increase significantly outpaces Moore's
Law—the observation that transistor density in integrated circuits doubles every two years, enabling
comparable increases in computing power, creating a growing gap between computing demands and
hardware efficiency improvements.
The carbon intensity of the electricity used for training also significantly impacts overall emissions. When
training is performed in regions with coal-heavy energy production, emissions can be up to 66 times higher
than in regions using primarily renewable energy sources. For instance, training a transformer model with
neural architecture search in the central United States produces 24 times the emissions as the same training
conducted in Quebec, which relies heavily on hydroelectric power [4].
45
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
What makes this consumption particularly concerning is its invisibility to end users. Unlike plastic waste
washing up on shores or vehicle emissions that can be seen and smelled, the environmental impact of
clicking a button to generate an AI response remains largely abstract and hidden from view. The
computational costs after deployment (inference) are lower per instance than training but rapidly
accumulate at scale, with an estimated 100 million users accessing popular AI services daily, even a modest
carbon footprint per query compounds into a significant environmental impact.
This invisibility problem is further exacerbated by the lack of standard reporting practices. Current research
practices rarely require disclosure of energy consumption or carbon emissions, making it difficult to track
or compare the environmental efficiency of different models or approaches. The research community has
begun advocating for standardized reporting of computational training time, energy consumption, and
carbon emissions, recognizing that what is measured is more likely to be optimized.
As AI capabilities continue to advance and become more deeply integrated into digital infrastructure,
addressing this invisibility challenge will be crucial for fostering more environmentally conscious
development and usage patterns, potentially through carbon-aware computing practices and greater
transparency regarding the environmental costs of AI systems.
Table 2: AI Training Environmental Impact [3, 4]
Training Scenario CO₂ Emissions (lbs)
GPT-3 626,000 lbs
NLP Model with Neural Architecture Search 626,000 lbs
Average American Car (Lifetime) 125,200 lbs (20% of GPT-3)
Training in a Coal-heavy Energy Region 66× baseline emissions
Training in Renewable Energy Region 1× baseline emissions
Central US Training (Transformer Model) 24× emissions compared to Quebec
Quebec Training (Transformer Model) Baseline for regional comparison
The Economic-Environmental Disconnect
Traditionally, economic and environmental costs have been somewhat aligned – resource-intensive
processes tend to be expensive, naturally imposing limits on their use. However, this relationship is
becoming increasingly decoupled in two critical ways:
When Cost Optimization Ignores Environmental Impact
As companies strive to reduce operational costs, they often focus on strategies that lower financial expenses
without necessarily addressing environmental concerns. Recent market analysis reveals a widening
disconnect between economic and environmental efficiency in cloud computing and AI services. Studies
measuring the carbon intensity of cloud instances found substantial variations depending on region and
time of day, with emissions differing by as much as 30× between the lowest and highest carbon intensity
46
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
timeframes [5]. This disparity demonstrates how cost optimization often fails to account for environmental
impacts.
Hardware optimization efforts primarily focus on performance improvements rather than environmental
efficiency. Data center carbon intensity varies dramatically by region, with values ranging from 210 to
1,060 gCO₂eq/kWh depending on the local electricity mix [5]. Despite this variation, pricing structures
rarely reflect these environmental costs, instead following market competition dynamics and regional
operating expenses. Research shows that even within the same cloud provider, instances in regions powered
primarily by fossil fuels are often priced 30-40% lower than those in regions with cleaner energy mixes [5].
This disconnect becomes particularly evident in hardware development strategies. Companies frequently
invest in making computational components like GPUs more cost-effective without proportionally reducing
material use or energy consumption. For instance, manufacturing processes that create cheaper GPUs while
still requiring similar quantities of silicon and rare earth materials optimize for economic rather than
environmental efficiency. Similarly, the practice of offshoring data centers to regions with lower
operational costs often prioritizes financial considerations over environmental impact. These regions may
offer reduced electricity prices and tax incentives but frequently rely on fossil fuel-heavy energy grids,
resulting in significantly higher carbon emissions per computation than facilities in regions with cleaner
energy infrastructures.
Data center relocation represents another dimension of this disconnect. Cloud providers frequently establish
operations in regions offering tax incentives and lower electricity rates, with minimal consideration for
energy sources. Efficiency improvements in computational systems often lead to increased total utilization
rather than reduced consumption—a phenomenon known as Jevons Paradox. A 6.8× improvement in
hardware efficiency over three years (2016-2019) corresponded with a 7.5× increase in computational
demand, essentially negating environmental benefits [6].
The Abstraction Gap
The second challenge stems from the "abstraction gap" – the cognitive distance between simple user actions
and their complex, resource-intensive consequences. The environmental impact of AI operations varies
dramatically depending on task complexity and implementation details. Training a large neural network
model produces CO₂ emissions that can be reduced by up to 99% by combining multiple efficiency
strategies, including algorithmic optimizations, hardware selection, and renewable energy sources [6]. For
example, simply switching from conventional GPUs to TPUs for large-scale model training can reduce
energy usage by 71% while maintaining computational output [6].
The geographic location of computation dramatically affects emissions, with the same workload potentially
producing anywhere from 30 to 840 gCO₂eq depending on where it's processed [5]. Carbon-aware
computing could significantly reduce environmental impact without compromising performance, yet
remains largely unimplemented in commercial systems.
47
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
This computational intensity remains invisible to users. A general language model inference request might
consume 0.07-0.05 kWh, seemingly small but significant when multiplied across billions of daily
interactions. The carbon footprint of cloud operations can be reduced by 34% simply by shifting
computation to times with carbon-efficient electricity availability [5].
In each case, minimal user effort initiates substantial computational work. This disconnect makes it
extraordinarily difficult for users to develop an intuitive sense of the environmental impact of their actions
without explicit feedback mechanisms or systems designed for carbon efficiency rather than merely cost
efficiency.
Table 3: Cloud Computing Carbon Intensity Variation [5, 6]
Region/Factor Carbon Intensity (gCO₂eq/kWh) Relative Price
Lowest Carbon Intensity Regions 210 100%
Highest Carbon Intensity Regions 1,060 60-70%
Lowest vs Highest Carbon Timeframes 1x vs 30x Same
Fossil Fuel-Powered Regions High 30-40% lower
Clean Energy Regions Low 30-40% higher
Case Study: The Viral #Ghibli Trend
The #Ghibli trend exemplifies the perfect abstraction gap problem identified earlier. When ChatGPT
introduced a feature allowing users to transform their photos into artwork reminiscent of Studio Ghibli's
distinct hand-drawn animation style, the environmental implications were completely hidden from user
view. Even technically sophisticated users who might understand the general computational intensity of AI
remained unaware of the specific resource demands of this feature. The transformation, triggered by a single
button click, concealed a process requiring substantial GPU resources—far more than typical text-based
interactions. The beautiful visual outputs, combined with minimal user effort, created a powerful incentive
for sharing and widespread adoption. This combination of hidden environmental cost and easy
reproducibility demonstrates how digital environments can rapidly accelerate environmentally damaging
behavior patterns without the visual feedback mechanisms that might otherwise constrain them. A major
AI provider introduced a feature allowing users to transform their photos into artwork reminiscent of Studio
Ghibli's distinct aesthetic. The beautiful results, coupled with the minimal effort required to create them,
sparked a viral sensation. The rapid adoption of this feature highlights how AI-enabled tools can quickly
scale across social networks, resulting in significant resource consumption. Research on viral AI feature
adoption shows that when consumer-facing generative AI features gain popularity on social media, usage
can spike by up to 500% within 48 hours, creating sudden and substantial environmental impacts before
mitigation measures can be implemented [7]. When applied to consumer-facing features that go viral, this
intensity effect is magnified exponentially.
48
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
The Amplification Effect of Virality
Social media's virality mechanisms create a particularly troubling dynamic when combined with resource-
intensive AI features. The diffusion of AI innovations follows distinct patterns from traditional
technologies, with adoption rates increasing by approximately 34% when visual outputs are easily shareable
on social platforms [7]. Organizations deploying AI features report a 42% increase in processing demands
when visual content generation capabilities are added to existing systems, particularly when these features
align with current aesthetic trends or cultural moments.
The environmental implications become apparent when examining the relationship between AI adoption
intensity and sustainability metrics. Studies indicate that high-intensity AI implementation without
corresponding environmental governance mechanisms increases the overall carbon footprint by 18-23%
[7]. This relationship is particularly pronounced in consumer-facing applications where usage patterns are
driven by social sharing rather than business necessity.
The #Ghibli trend exemplifies how social media accelerates resource-intensive AI feature adoption.
Research on social media marketing shows that AI-generated visual content receives 3.7 times more
engagement than non-AI content, creating powerful incentives for widespread use despite environmental
costs [8]. The phenomenon is further reinforced by the fact that 71% of social media users are more likely
to share AI-transformed personal content compared to other media types.
Technical and Environmental Consequences
The consequences were immediate and severe. The provider reported significant strain on their GPU
infrastructure, ultimately forcing them to implement rate limiting on the feature. While the primary concern
was likely financial – GPU compute time translates directly to operational costs – the incident perfectly
encapsulates our environmental challenge.
Studies on AI in social media marketing reveal that image transformation features consume approximately
0.3-0.5 kWh per image processed, with viral trends potentially generating millions of transformations daily
[8]. When content goes viral, platforms experience what researchers term "adoption cascade effects," where
feature usage can increase by 300-500% within 24-48 hours. During such periods, environmental
considerations are rarely prioritized, with only 12% of platforms having automatic sustainability-focused
throttling mechanisms.
For users, the experience was frictionless: upload a photo, click a button, share the result. The underlying
reality – thousands of processors running at maximum capacity, consuming megawatts of electricity –
remained entirely hidden from view. Research indicates that transparency regarding AI's environmental
impact significantly affects user behavior, with 63% of users reporting they would moderate their usage if
provided real-time environmental impact data [8]. However, only 8% of platforms currently provide such
information.
49
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
This pattern can create environmental impact spikes that occur faster than companies can implement
reasonable constraints. By the time a company identifies the problem and implements rate limiting, millions
of unnecessary compute operations may have already occurred, creating significant environmental costs
that remain largely invisible to end users.
Table 4: Viral AI Feature Adoption Patterns [7, 8]
Factor Measurement
High-intensity AI Resource Utilization Increase 27.30%
Visual AI Adoption Rate Increase 34%
Visual Content Processing Demand Increase 42%
Carbon Footprint Increase (High-intensity AI) 18-23%
AI-generated Visual Content Engagement 3.7x
AI-transformed Personal Content Sharing Likelihood 71%
Energy per Image Transformation 0.3-0.5 kWh
Platforms with Sustainability Throttling 12%
Users Who Would Moderate Usage with Impact Data 63%
Platforms Providing Environmental Impact Information 8%
Solutions and Future Directions
Addressing the full lifecycle of AI's environmental impact requires different strategies for both training and
inference phases. While training optimization focuses on one-time efficiency improvements and
appropriate model sizing, inference optimization demands ongoing efficiency in serving billions of daily
interactions. Carbon-aware inference systems could intelligently route requests to data centers with the
lowest current carbon intensity, potentially reducing emissions by 45-60% compared to static routing
approaches. Additionally, implementing differential service levels based on computational intensity and
environmental impact—offering standard response times for efficient requests while queuing resource-
intensive operations during periods of renewable energy abundance—could balance user experience with
environmental responsibility.
Addressing this challenge requires a multi-faceted approach that bridges the visibility gap between user
actions and environmental consequences. Recent frameworks for sustainable AI demonstrate that
systematic implementation of mitigation strategies could reduce carbon emissions by up to 70% while
maintaining comparable service quality [9].
Environmental Impact Transparency
AI service providers should consider implementing environmental impact indicators that make resource
consumption visible to users. Recent studies on sustainable computing practices show that transparency
tools lead to significant behavioral changes. When energy consumption dashboards were implemented in
enterprise settings, organizations reduced their AI-related energy usage by 23.6% over a six-month period
50
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
[9]. This reduction was attributed to increased awareness and operational adjustments rather than decreased
functionality.
Carbon footprint calculators for AI workloads have proven particularly effective, with accuracy rates now
reaching 94.2% for common enterprise applications [9]. User studies reveal that presentation format
significantly impacts decision-making, with visual indicators of computational intensity resulting in 38.7%
more efficient resource allocation compared to numerical representations alone.
Environmentally-Conscious Rate Limiting
Rather than implementing rate limits solely based on economic considerations, providers could establish
environmentally motivated usage caps. Research on sustainability-focused computing frameworks
indicates that carbon-aware scheduling can reduce greenhouse gas emissions by 15-45% without degrading
service availability [10]. These approaches provide a balanced solution that maintains accessibility while
reducing environmental harm.
Dynamic pricing strategies that incorporate environmental costs have shown promising results in field tests.
When carbon intensity was incorporated into pricing models, non-essential computational tasks naturally
shifted to periods of renewable energy abundance, reducing carbon footprint by 32.4% while maintaining
overall productivity [10]. Time-based access tiers and environmental quotas have demonstrated similar
effectiveness in enterprise deployments.
Architectural Innovation for Sustainability
The AI industry needs to prioritize architectural approaches that reduce environmental impact. Recent
advances in model compression techniques have achieved 80% reductions in computational requirements
with only 2.7% performance degradation on standard benchmarks [9]. These efficiency gains directly
translate to reduced energy consumption and environmental impact.
Technical research on specialized computing architectures shows that application-specific hardware
designs improve energy efficiency by factors of 3-7× compared to general-purpose computing platforms.
Integration of renewable energy into data center operations now provides tangible benefits, with smart load
balancing systems reducing carbon intensity by 51.3% while maintaining 99.98% service reliability [10].
User Education and Feedback Mechanisms
Perhaps most importantly, effective user education approaches are critical for sustainable AI adoption.
Studies on human-computer interaction show that contextualizing environmental impact in familiar terms
increases user engagement with sustainable practices by 47.2% [9]. Interactive tools that visualize abstract
concepts like energy consumption have proven particularly effective.
Comparative feedback mechanisms that benchmark usage against industry averages or organizational goals
motivate sustainable behavior, with documented reductions of 25.8% in unnecessary computational
requests. Gamification elements further enhance engagement, with recognition systems for
51
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
environmentally-conscious practices increasing consistent application of sustainable computing behaviors
by 32.7% [10].
Fig. 1: Sustainable AI Implementation Benefits [9, 10]
CONCLUSION
The integration of generative artificial intelligence into everyday digital experiences has created a
paradoxical relationship between technological advancement and environmental stewardship. As
conversational interfaces, voice technologies, and image generation capabilities transform how people
interact with technology, the hidden environmental costs continue to accumulate in ways that remain largely
invisible to end users. The abstraction gap between simple actions and their resource-intensive
computational consequences presents a fundamental challenge to developing environmentally conscious
usage patterns. This challenge intensifies when viral adoption of features like the #Ghibli transformation
rapidly scales resource consumption before appropriate constraints can be implemented. The economic-
environmental disconnect further complicates matters as optimization strategies that reduce financial costs
frequently fail to address—and sometimes exacerbate—environmental impacts. Looking forward, bridging
this visibility gap represents the most promising pathway toward sustainable AI deployment. Implementing
transparent impact indicators, environmentally-conscious rate limiting, efficient architectural innovations,
52
European Journal of Computer Science and Information Technology,13(33),43-53, 2025
Print ISSN: 2054-0957 (Print)
Online ISSN: 2054-0965 (Online)
Website: https://www.eajournals.org/
Publication of the European Centre for Research Training and Development -UK
and effective user education creates opportunities to align technological progress with environmental
responsibility. The integration of carbon-aware computing practices, feedback mechanisms that encourage
mindful usage, and hardware designed for sustainability rather than raw performance offers substantial
potential for improvement. Only by recognizing and addressing these hidden environmental costs can
generative AI fulfill its transformative potential without undermining the natural systems upon which
humanity depends. This requires intentional collaboration across stakeholders to ensure that impressive
technical capabilities evolve in parallel with robust environmental governance frameworks.
REFERENCES
1. Itransition, "The ultimate list of conversational AI statistics for 2025," 2025. [Online]. Available:
https://www.itransition.com/ai/conversational
2. Dasha, "The Role of Voice AI in Environmental Sustainability and Sales," 2024. [Online].
Available: https://dasha.ai/blog/the-role-of-voice-ai-in-environmental-sustainability-and-sales
3. Peter Henderson, et al., "Towards the Systematic Reporting of the Energy and Carbon Footprints
of Machine Learning," arXiv, 2022. [Online]. Available: https://arxiv.org/abs/2002.05651
4. Emma Strubell, Ananya Ganesh, and Andrew McCallum, "Energy and Policy Considerations for
Modern Deep Learning Research," Proceedings of the AAAI Conference on Artificial
Intelligence, 2020. [Online]. Available: https://ojs.aaai.org/index.php/AAAI/article/view/7123
5. Jesse Dodge, et al., "Measuring the Carbon Intensity of AI in Cloud Instances," ACM Digital
Library, 2022. [Online]. Available: https://dl.acm.org/doi/10.1145/3531146.3533234
6. David Patterson, et al., "Carbon Emissions and Large Neural Network Training," arXiv, 2021.
[Online]. Available: https://arxiv.org/abs/2104.10350
7. Jiachen Li and Xiu Jin, "The Impact of Artificial Intelligence Adoption Intensity on Corporate
Sustainability Performance: The Moderated Mediation Effect of Organizational Change,"
ResearchGate, 2024. [Online]. Available:
https://www.researchgate.net/publication/385337595_The_Impact_of_Artificial_Intelligence_Ad
option_Intensity_on_Corporate_Sustainability_Performance_The_Moderated_Mediation_Effect_
of_Organizational_Change
8. Ms. Anshu and Monika Sharma, "AI in Social Media Marketing: Opportunities and Challenges,"
ResearchGate, 2024. [Online]. Available:
https://www.researchgate.net/publication/384579089_AI_in_Social_Media_Marketing_Opportun
ities_and_Challenges
9. Vivian Liu and Yiqiao Yin, "Green AI: exploring carbon footprints, mitigation strategies, and
trade-offs in large language model training," Discover Artificial Intelligence, 2024. [Online].
Available: https://link.springer.com/article/10.1007/s44163-024-00149-w
10. Saumya Dash, "Green AI: Enhancing Sustainability and Energy Efficiency in AI-Integrated
Enterprise Systems," IEEE Xplore, 2025. [Online]. Available:
https://ieeexplore.ieee.org/document/10849555
53
