1 Introduction

The accelerating development and deployment of artificial intelligence (AI) has sparked intense ethical debate, not only concerning fairness, accountability, and transparency, but increasingly about its environmental impacts. This growing debate about the environmental implications of AI, related to energy consumption [76, 112] or water consumption [72, 73] highlights that the use of AI systems in many domains comes at a cost. While AI is often portrayed as a tool to mitigate climate change and optimize resource management, scholars are now turning attention to the sustainability of AI itself, namely its carbon footprint, water footprint and other material dependencies [18, 125] pointing on an extractivist blindspot, [55, 128] that neglects the material constitution of such technologies. The expansion of digital infrastructures, especially the building of new hyperscale data centers but also the need for more hardware, such as GPU, TPU, drones, sensors, and lidar is always associated with environmental consumption. Amid the global climate crisis and accelerating biodiversity loss, AI’s promised benefits are increasingly questioned in light of its infrastructural prerequisites and their ecological costs. The rhetoric of AI futurism abstracts technology from its material impacts and reinforces a progress narrative that sidelines planetary futures [101]. In response, there are growing calls for an environmental-ethical and global-justice orientation of AI that extends beyond distributive concerns to include recognition justice and the intrinsic value of ecosystems and non-human life [126] as well as concepts of planetary justice [111]. These efforts show how AI relies on energy-intensive infrastructures, extractive supply chains, and unequal labor conditions that undermine ecological integrity and human flourishing, particularly in the Global South (Sahebi and Formosa 2023, Downey 2024). Rather than a universally emancipatory project, AI is increasingly understood as embedded in imperial extractivist logics, intensified by the concentration of technological power in a few corporations [49, 115]. To better understand how the materialities and environmental impacts that occur throughout the AI-Lifecycle are related to vulnerabilities, inequality, and more-than-human worlds, we argue that two things are key. First, to define AI as a socio-technical-ecological system, and second, to take the lens of different justice perspectives. Such perspectives highlight how AI systems not only consume energy and resources but also shape human and non-human relations in ways that reproduce or exacerbate inequalities. In this paper, we aim to analyse the environmental consequences of AI from three different justice concepts,namely Environmental Justice (EJ), Climate Justice (CJ) and Multispecies Justices (MSJ). The aim of this paper is to apply three complementary justice perspectives to the environmental impacts associated with the infrastructural prerequisites of the AI lifecycle. AI systems cannot operate without these infrastructural prerequisites because their training, deployment, and maintenance depend on material and energy-intensive systems. Production requires specialized hardware and raw materials, operation relies on data centers, electricity grids, and network infrastructures, and disposal depends on waste and recycling systems.

The remainder of the paper is organized as follows: The next section will form the foundation of the paper by formulating a conceptualization of AI as a socio-technical-ecological system. In the third section we will have a deeper look at the diverse environmental impacts in the production, operation and disposal phases of AI-related infrastructures. In the fourth section, we explain the three different justice perspectives of Environmental Justice, Climate Justice, and Multispecies Justice and illustrate what they make us see, when looking at the environmental impacts that derive from the infrastructural prerequisites of of AI-Systems. In Section five, we illustrate how these three perspectives provide us with reflective lenses for understanding how the materialities and environmental lifecycle impacts of AI shape human and non-human relations and in which ways they reproduce or exacerbate inequalities, based on different aspects of the AI-lifecycle. We conclude by discussing approaches on how to reduce those impacts.

2 AI as an embedded system and its infrastructural prerequisites

There is a predominant view of looking at AI as an abstraction (or an isolated technical object with no connection to its surrounding). In this view, AI is seen as software, algorithms, and models which are assigned with tasks we relate to intelligence. This is also called the abstraction view of AI [47, 102]. This approach to AI systems, which we will call the artefact-only approach, is insufficient, as critics argue for decolonial ethical frameworks, that restructure power relations and center the most vulnerable, ultimately requiring AI to be understood as a socio-technical system to adequately address issues of fairness, privacy, and responsibility [32, 80, 102]. The narrow artefact-only approach treats AI systems as a self-contained technological system. Such an approach is insufficient because questions of fairness, privacy, and responsibility require understanding AI as a socio-technical system rather than a standalone technical object [32, 102], while it also sidelines the environmental and infrastructural conditions that make AI systems possible, thereby obscuring their dependence on global networks of computation, resource extraction, and energy and water consumption. Consequently, current scholarship calls for a holistic conceptualization of AI, which situates its functionality and effectiveness within the infrastructural prerequisites and the socio-technical and ecological systems on which it relies [55, 66, 96] grounded in socio-technical systems thinking [65]. In order to meet the demand for integrating ethical considerations early in the technological development process [6, 7, 34] there is a need to situate AI within broader ecological, political, and social contexts rather than treating it as a neutral or isolated technology [18]. This aligns with many of the works mentioned earlier [126, 134], as well as other advocates that argue that AI systems should be approached as a technology embedded in its socio-technical context [47]. This embedded view highlights that AI is never purely technical: it is historically layered, socially organized, and politically structured and woven into routines, institutions, and material systems.

Reflecting on the mentioned criticisms of the narrow definition and in consideration of the mentioned complex interaction of AI systems with their socio-environmental context, we base our definition on a broader understanding of AI systems as socio-technical-ecological systems [96] because if AI is only treated as a socio-technical system, climate impacts, extractivism, and ecological harm may remain invisible, which exacerbates the extractivist blind spot [55, 128]. This perspective emphasizes that AI does not exist in isolation but is dependent on extensive infrastructural prerequisites such as energy grids, data centers, and global supply chains. As socio-technical–ecological systems, AI systems are understood within broader ecological, political, and social contexts, which we refer to as the contextualized view. Enhancing existing scholarship, this approach reframes AI from an isolated technology into a broader system that encompasses technical, material, social, and environmental dimensions.

From this contextualized view, it becomes clear that AI systems are intertwined with infrastructures that emerge through practices and relational entanglements [109]. Infrastructure is embedded in routines, practices, and institutions, and is often operating unnoticed until a failure or breakdown makes it visible. It has a wide reach or scope, crossing multiple contexts and domains, and it requires skills to use it effectively and is always built on an installed base, relying on legacy systems, standards, and protocols that structure how new technologies are integrated [109]. Once those relations made up of high performance computing hardware production, hyperscale data centres, and data transmission network maintenance are established, they generate dependencies that are difficult to reverse. Acknowledging these infrastructural prerequisites reveals that AI systems privilege efficiency over ecological resilience, locking societies into resource-intensive, unsustainable pathways that favor incremental fixes over structural change [95].

3 Impacts across the AI lifecycle: production, operation, disposal

AI-systems depend on a vast diversity of infrastructural prerequisites and within the last years there is growing evidence about their impact on the environment, with issues such as computing related carbon footprint [77, 112], as well as energy [76] and water consumption of hyperscale data centres that enable large-scale model training and inference [62, 72, 73]. More recently, mining [88] and landuse [81] have also become prominent issues. Building on this argument, this section adopts a perspective on AI-systems that enhances the more commonly recognized stages of AI design, development, and deployment with the lifecycle-phases related to its infrastructural prerequisites, namely, production, operation, and disposal (see Fig. 1). Given that architectural choices and design goals such as performance, scalability, and accuracy determine how AI systems are built and run, they also shape the infrastructural prerequisites of AI across its lifecycle, including patterns of production, operation, and disposal. Decisions made during design and development influence hardware requirements, storage and networking needs, operational demands such as inference frequency and cooling, and even how easily systems can be repaired or recycled. At the same time, access to high-performance computing concentrates workloads into energy- and water-intensive peaks, while supporting activities like data storage and logging generate a continuous background load, showing how design, development, and deployment are tightly coupled to the material and environmental conditions that sustain AI systems.

Fig. 1
Fig. 1
Full size image

The infrastructural prerequisites of AI-Lifecycles (own figure)

From this contextualized perspective, the AI lifecycle is understood as embedded within and dependent on extractive supply chains, energy and water infrastructures, and waste and recycling systems [48] that shape how AI is produced, operated, and disposed of. Accordingly, this section focuses on the environmental impacts that arise across these interconnected phases of the AI lifecycle.

3.1 Environmental impacts of production: mining, hardware, chip production

AI production involves multiple processes and has wide-ranging environmental impacts. In what follows, we only examine mining and semiconductor manufacturing as key drivers of the production footprint.

3.1.1 Mining for raw materials and rare minerals

The supply chains underlying AI hardware also impose substantial environmental burdens because of the raw materials and rare minerals that are needed for the manufacturing of CPU and GPU. The production of critical minerals such as lithium, cobalt, graphite, and copper is projected to rise sharply by 2040 [59]. The rare earth elements and transition minerals mentioned above are essential not only for AI hardware but also for low-carbon energy technologies, creating competing pressures between digitalisation and climate transition goals [59]. As demand rises, ore grades decline, meaning larger areas must be mined to produce the same output. Greenpeace [43] further notes that as ore grades decline, the extraction of critical raw materials becomes increasingly energy- and water-intensive, deepening the environmental footprint of AI hardware.

The materials are frequently extracted in environmentally sensitive areas under conditions that generate deforestation, soil degradation, and water pollution, [64, 82]. The geographical concentration of extraction, which accounts for more than half of global cobalt production in the Democratic Republic of Congo and a major amount of lithium in Australia and Chile, creates vulnerabilities for both ecosystems and local communities, often with little corresponding benefit from AI development [13, 93, 122]. The Greenpeace [43] report characterizes these processes as a form of “digital colonialism,” where communities in the Global South, especially in cobalt-, lithium-, and nickel-mining regions, bear the brunt of deforestation, water depletion, and toxic exposure. This dynamic is especially paradoxical because many of the same AI systems driving these extractive issues are simultaneously promoted as solutions to environmental degradation, through applications such as simulations of restoration scenarios or predictive analytics of wildfires [84]. An example of the environmental toll of mining materials used in AI systems comes from copper extraction in Eastern Serbia. Adamovic et al. [3] show that acid mine drainage (AMD) from the Bor copper district produces highly acidic runoff carrying heavy-metal loads of iron, copper, and arsenic, contaminating river systems more than 100 km downstream. These findings illustrate how the extraction of AI-related critical materials can generate long-lasting ecological harm through water acidification and heavy-metal mobilization. Historical analyses reveal that these patterns are not new. Acosta, Purbrick and Ribas [2] trace how nitrate and copper mining in Chile’s Atacama Desert turned the region into a “satellite of global capitalism,” supplying material wealth to industrial centres in Europe while leaving behind waste, pollution, and social dispossession. Moreover, research in Indonesia shows that deforestation in nickel-mining regions nearly doubled between 2011 and 2018, when global demand surged [75]. Similarly, the environmental footprint of copper production in China, measured as the land needed to provide resources and absorb emissions, is expected to more than double from 2010 to 2050 [29]. These examples underscore that the environmental consequences of the production phase are not abstract projections but measurable realities.

3.1.2 Semiconductor manufacturing

An important and often neglected impact according to Greenpeace East Asia [42], is the global electricity consumption associated with AI-related chip production, which grew by more than 350 percent between 2023 and 2024, while related emissions from electricity use increased by 357 percent over the same period. Projections suggest that by 2030, electricity demand from AI chipmaking could reach levels up to 170 times higher than in 2023, exceeding the current annual electricity consumption of some European countries. This trajectory is particularly significant for East Asia, where a large share of global semiconductor production is located, and where electricity systems remain heavily dependent on fossil fuels. In 2023, fossil fuels accounted for 58.5 percent of electricity generation in South Korea, 68.6 percent in Japan, and 83.1 percent in Taiwan [42]. Chip production is also highly water-intensive, with a single advanced fabrication plant consuming tens of millions of litres of water per day to produce ultrapure water for manufacturing processes, placing additional pressure on local water systems [53]. Furthermore, transport of raw materials and intermediate components across global supply chains, as well as hazardous on-site waste from chemical processing, add to the upstream ecological burden of semiconductor manufacturing [53].

Moreover, new fossil fuel infrastructure is being developed to accommodate the anticipated rise in demand. In 2024, the South Korean government approved the construction of a 1 gigawatt liquefied natural gas power plant to supply electricity to SK Hynix’s Yongin Semiconductor Cluster [42]. In addition, plans are underway to build another three gigawatts of liquefied natural gas power capacity to meet the energy needs of Samsung’s National Semiconductor Cluster in the same region [41]. What these developments also show is that the rising energy demands related to hardware production for making AI systems work, shape energy policy agendas and may lead to trajectories that even increase carbon emissions.

3.2 Environmental impacts of operation: energy, water and geographic footprint of server infrastucture

The AI operation stage involves multiple processes and has wide-ranging environmental impacts. Central to this phase are hyperscale data centres, which represent a growing trend in global digital infrastructure [43]. Unlike conventional data centres, hyperscale facilities are designed for extremely high computational density and capacity, making them necessary for large-scale model training and inference [43]. In what follows, we only examine computing-related energy use, water footprint, and land use and habitat loss.

3.2.1 Computing-related energy use and carbon footprint

At the beginning of this debate around 2019 the energy consumption and associated carbon emissions that derive from the development (i.e. Neural Architecture Search, NAS) and training of Large Language Models (LLMs), was mentioned as a critical issue [112], since training runs of large models are particularly energy-intensive. For example, the final training of BLOOM, a 176-billion-parameter model, generated 24.7 metric tons of CO2-equivalent emissions, despite being conducted in France, where electricity production is relatively low-carbon due to nuclear power, [59, 77]. By contrast, comparable models trained in regions with more carbon-intensive electricity sources produced far greater emissions: GPT-3 (175B) training emitted 552 tons CO2eq, Gopher (280B) 380 tons, and OPT-175B approximately 76 tons [77]. However, there is no publicly available information about the energy consumption or carbon footprint of the most recent large-scale models such as GPT-4 or GPT-5, which are likely to be considerably higher given their increased model size and computational demands. Yet while training got attention for its extraordinary one-time costs, evidence increasingly indicates that inference is the more consequential long-term driver of emissions. Therefore, while the emissions from a single inference are small, when scaled across millions of users and requests, inference can account for a larger share of the total lifecycle emissions than training, depending on model size, efficiency, and hardware [76]. Furthermore, hardware choices significantly affect energy consumption. For instance, GPUs can deliver much higher performance for parallel workloads, but also draw substantially more power during runtime than CPUs with comparable tasks, especially under heavy load and in large clusters [22]. Data-pipeline impacts grow when data sets are large and poorly managed, increasing embodied and operational emissions, thereby intensifying the environmental externalities associated with data acquisition, curation, storage, and transmission across the AI life cycle.

3.2.2 Water footprint

Alongside carbon emissions and electricity demand, water consumption has emerged as a crucial dimension of AI’s environmental footprint. As mentioned before, modern AI systems depend on hyperscale data centres that require vast quantities of water, primarily for cooling. As AI adoption accelerates globally, the trend points to an escalating and geographically uneven water burden. A consistent pattern across regions is that the expansion of AI systems drives a steep rise in water withdrawals. Data centres in hot or water-stressed areas consume millions of litres daily, and the cumulative effect is striking: training a single GPT-3 model in Microsoft’s U.S. facilities directly evaporates around 700,000 litres of freshwater [72, 73]. At scale, global AI demand is projected to account for 4.2–6.6 billion cubic meters of water withdrawal by 2027, surpassing the annual water use of half the United Kingdom [72, 73]. This growth reflects not just one-off training events but also the continuous demand of inference services, which must keep servers running and cooled around the clock.

Corporate reports already reveal the consequences of this trend. Microsoft’s water consumption has risen by 34% and Google’s by 20%, largely due to AI-driven data-centre operations [72, 73]. Such increases illustrate how quickly AI can shift utility demands even for companies that operate some of the most advanced and efficient infrastructures. At the same time, these companies increasingly position AI as a tool for addressing water scarcity and inefficiency, for instance through AI water management systems that deploy generative AI to detect leaks, reduce wastage, and optimize water use [133]. On the user side, the water footprint translates into surprisingly tangible terms: generating just 10–50 medium-length responses with GPT-3 is equivalent to consuming a standard 500 ml bottle of water, depending on deployment location and cooling technology [72, 73].

Taken together, these findings suggest a structural trajectory: as AI models grow larger and their applications more pervasive, water demand from supporting infrastructure will climb sharply. This pressure will be felt most acutely in regions already facing water stress [123], but even in relatively water-rich areas, AI’s “thirst” is becoming a notable part of the environmental debate. The challenge ahead is not only to decarbonize AI but also to make it less water-intensive through innovations in cooling, siting, and system design.

3.2.3 Landuse & habitat loss

At the infrastructure level, the expansion of hyperscale data centres, that cover areas of up to four square kilometers is transforming landscapes in both urban and rural settings [43]. Data centre projects often require thousands of acres of flat land with access to energy and transmission lines. Recent developments in the United States illustrate this shift: plans for 37 data centres on 1700 acres in Virginia sparked legal disputes over their proximity to historic conservation sites [51], while Meta purchased 2700 acres of farmland in the Mississippi River basin to construct its largest-ever data centre [27].

Furthermore, data centers require extensive land not only for their buildings but also for energy infrastructure, cooling systems, and supporting logistics. Large-scale projects often convert agricultural land or undeveloped ecosystems into industrial landscapes, contributing to habitat fragmentation, biodiversity loss, and urban sprawl [40, 103]. Renewable energy projects designed to offset data center emissions, such as solar farms or wind parks, may themselves disrupt habitats when poorly sited. Taken together, these dynamics underscore that AI’s land-use impacts are twofold: they are embedded in the extraction of non-renewable resources that underpin hardware, and they materialize in the direct conversion of land for data centre facilities. As AI systems expand, these pressures will increasingly reshape landscapes and intensify conflicts over land between energy transition, food production, heritage conservation, and digital infrastructure.

3.3 Environmental impacts of disposal: increase of hardware use and electronic waste handling

Electronic waste (e-waste) is commonly defined as discarded electrical or electronic equipment and its components, whether destined for reuse, recycling, recovery, or final disposal. It represents a dual challenge, since e-waste streams contain hazardous materials such as lead, mercury, and brominated flame retardants, while also holding valuable secondary resources including gold, silver, and palladium [38]. The rapid growth of AI systems has a widespread impact on the global e-waste distribution by adding a large, geographically concentrated waste stream to an already expanding sector. In particular, the training of generative AI models relies on GPU-enabled servers, which are deployed at scale to process vast quantities of data but often become redundant once training is complete. From an e-waste perspective, this creates a risk of short hardware lifetimes, as GPUs and entire servers may still be technically functional yet rendered obsolete by lack of sustained demand [119]. Again on the contrary, there is a growing promotion of AI solutions for e.g. waste management [33] and material-efficient product design [12], revealing a tension between AI-based optimization and the AI-specific hardware entering the global e-waste stream.

Estimates suggest that generative AI alone could generate 1.2–5.0 Mt of cumulative waste by 2030, raising global e-waste totals by 3–12% as annual volumes surge from 4.6 (2022) to 43.2 Mt (2030). Here, e-waste refers specifically to end-of-service flows from LLM servers and their core components: GPUs, CPUs, memory, storage, communication modules, and power systems [131]. The scale is intensified by the sheer size of components, for example one Nvidia Blackwell server exceeds 1.3 t, and by the hazardous mix of materials: printed circuit boards, batteries, and structural parts contain lead, chromium, and other heavy metals, alongside recoverable precious metals such as gold, silver, and palladium. Yet recovery remains limited: a Greenpeace report [43] notes that low recycling rates for chips and critical raw materials hinder circularity, even as total annual server/storage waste grows from 0.2 (2023) to 0.8 Mt (2030).

Compounding risks include the illegal export of e-waste, documented for consumer electronics such as laptops, monitors, and printers. Around 6% of tracked items are exported from Europe and about 40% from the United States, reflecting a broader trend: in 2022, an estimated 5.1 billion kg of e-waste crossed borders, with 65% moving from high- to middle- and low-income countries through uncontrolled routes [38]. These unmanaged flows create uncertainty over how AI-specific hardware will be treated at end of life. Schneider et al. [99] assess Google’s Tensor Processing Units, finding retirement emissions small but likely to grow, highlighting the need for tighter monitoring and circular disposal systems. Although recycling prevents the extraction of 900 billion kg of ore and recovers USD 28 billion in metals annually, only 22% of global e-waste is formally processed [43]. Recycling remains limited as critical materials and circuit boards are rarely recovered, while informal dismantling and burning perpetuate pollution and health hazards [25]. E-waste contamination contributes to respiratory illness, cancer, reproductive harm, and widespread water pollution in countries like India, intensified by illegal waste imports [17].

4 Three justice perspectives: approaches on environmental impacts of AI systems

In this section, we outline the three justice perspectives that are used to examine the infrastructural prerequisites along the AI-Lifycycle: environmental justice, climate justice, and multispecies justice. These perspectives have emerged at different moments and address distinct yet overlapping dimensions of socio-environmental issues. Environmental justice provides a foundation for understanding unequal distributions of environmental harms, climate justice extends this focus to climate-related and intergenerational dimensions, and multispecies justice foregrounds the moral and ecological significance of non-human life. The three frameworks are not presented as hierarchically ordered or as competing approaches, but as complementary lenses for examining the environmental and socio-ecological challenges associated with AI lifecycles.

4.1 Environmental justice

The term Environmental Justice (EJ) can be understood both as a political demand and as an interdisciplinary field of research and analytical perspective for investigating ecological distribution conflicts [52]. Environmental justice studies emerged as an interdisciplinary body of literature, in which researchers were documenting the unequal impacts of environmental pollution on different social classes and racial/ethnic groups [79]. Early Environmental Justice activism can be tied back to the 1980s, built on the civil-rights movement, the labour movement, indigenous rights, and anti-toxic waste movements in the U.S.A, with key policy and institutional steps occurring in the early 1990s. Environmental justice arises from the recognition that environmental burdens and benefits are distributed unevenly, often along lines of race, class, indigeneity and political power, and it demands that these inequities be treated not as incidental outcomes but as matters of structural injustice requiring ethical and political redress [132]. At its core, the concept challenges the assumption that all communities experience environmental risks equally, drawing attention to how toxic exposures, resource extraction, inadequate infrastructure and climate vulnerabilities disproportionately affect marginalised populations [114]. Meta-analyses confirmed that race, often more than income, is a strong predictor of proximity to environmental hazards, with disparities extending beyond siting to include slower government cleanups and weaker enforcement in affected areas [20, 79, 94]. Explanations for these inequities emphasize economic inequality, sociopolitical exclusion, and racial discrimination. Poor communities often lack the political power to resist hazardous location decisions, while racial prejudice and structural inequalities further compound environmental risks. A “chicken-and-egg” debate persists over whether industries deliberately target minority neighborhoods or whether marginalized groups settle in polluted areas due to cheaper housing, however, evidence suggests that intentional targeting has frequently occurred [79].

While these studies primarily highlight the unequal distribution of environmental burdens, what is mainly referred to as distributive injustice, Environmental Justice as a concept has evolved to encompass broader dimensions. Rather than focusing solely on distributive fairness, Environmental Justice also includes procedural and recognition dimensions. Procedural justice emphasizes the fairness and inclusivity of decision-making processes, ensuring that affected communities have genuine opportunities to participate, influence outcomes, and access transparent information. Recognition justice, on the other hand, calls for the acknowledgment and respect of marginalized groups’ histories, cultural identities, and knowledge systems, affirming them as legitimate foundations for environmental policy and advocacy [35]. As Figueroa [35] argues, without recognition, participation often becomes tokenistic, as the values and worldviews of dominant groups continue to define what counts as valid knowledge or concern. Likewise, procedural justice alone cannot address the deeper structural and cultural exclusions that prevent some communities from being heard or respected in the first place; recognition justice seeks to repair those underlying injustices by affirming the moral and epistemic standing of all communities.The framework has evolved to include concerns about intergenerational responsibility, the legacies of colonialism and the intersectional nature of oppression, bridging environmental ethics with civil rights, public health and decolonial thought. In doing so, it not only critiques dominant models of environmental governance but also reorients environmentalism towards equity, accountability and community self-determination, positioning justice as a necessary condition of sustainability rather than a secondary consideration. Concepts such as “discriminatory environmentalism” highlight how lack of opportunity for democratic participation in the environmental movement exacerbates these inequities [36]. Environmental Justice concerns operate both on a domestic level, affecting people of color and poor populations, and global level; impacting communities in the industrially marginalized Global South [36]. Over the past twenty years, climate change has become a central concern within Environmental Justice, broadening the discussion to encompass global responsibility.

4.2 Climate justice

The theory of Climate Justice began emerging in the 1990s after a variety of social and Environmental Justice movements were active in opposition to multiple fossil fuel industry operations [19], and grew more prominent after the shortcomings of global climate governance model in COP15 in 2009 [60, 98]. Many tie the origins back to the Environmental Justice and civil society movements in the late twentieth century [4, 113], while others attribute its rise to the anti-colonial indigenous communities of the global south [37, 87].

Climate Justice encompasses many different notions of how justice is defined, and there have been different approaches to outline these different notions. Distributional justice investigates questions related to even or equal distributions of aspects related to climate change. These aspects can be in the form of impact, contribution, vulnerability, responsibility, suffering, etc., and across different varying factors such as region, wealth, gender, caste, ethnicity, and class, and can also be cross-generational. There are also frameworks for identifying aspects of distribution, one notable example is to identify three aspects; the goods and ills being distributed, to whom or what it is [not] being distributed, and which is the most appropriate mode of distribution [78, 107, 108]. Procedural justice engages with questions regarding fair and just participation, representation, involvement, and engagement on an international, national, regional, and local level related to the climate change-related theme under focus. These levels are subject to different definitions of what constitutes fair representation and participation. Some varying definitions of participation here include; representation in governance [91], meaningful involvement [106], involvement in decision-making and research [135]. But apart from questions of participation and representation, some broaden the definition beyond political engagement and representation. Zimm et al. [135] argues that procedural justice should also be extended to modelled future scenarios (in terms of choice of model selection in the form of transparency about underlying assumptions), while Heffron [50] argues that it involves provision of legitimate opportunities in all stages of a project, including planning, construction, operation, and end-use. Corrective justice, also referred to as restorative justice [120], focuses on questions regarding restoration, compensation, responses, and rectifications, and what they would look like in contextual form Apart from these three dimensions, other authors argue for different definitions involving other dimensions of Climate Justice. Amongst those are the notions of transitional justice; which considers questions related to achieving a just transition in the context of climate change (i.e. just transition to a carbon–neutral market) [61], and recognitional justice; those affected by Climate Change are represented in formal equity with considerations of the differences in values and perspectives [56].

Climate Justice as a framework has also been criticized by other justice theories. One prominent criticism is by Environmental Justice, indicating that looking at the issue from climate considerations perspective engages us more with intergenerationally, as these impacts are more future-oriented, and less likely to engage with immediate environmental hazards [89]. Other criticisms also include Critical Climate Justice theories, which advocate more systemic changes to address inequalities and demand less obscurity on issues such as structural domination, colonial histories, and political struggles (Sultana 2021), and Multispecies Justice (See Sect. 4.3).

4.3 Multispecies justice

Multispecies Justice has emerged as a critical reorientation within environmental and political theory, responding to the inadequacies of prevailing frameworks that continue to privilege human-centered concerns and thereby perpetuate anthropocentrism. It developed out of the criticism that Climate Justice is insufficient to overcome the persistent silencing of voices and that it implicitly condones human exceptionalism [121]. Therefore, this approach seeks to reconceptualize Climate Justice in a way that is more inclusive, acknowledging the diverse histories and experiences of social, environmental, and ecological harm, while simultaneously opening just pathways toward uncertain futures.

At its core, Multispecies Justice rejects the triad of assumptions that underpin human exceptionalism: first, that humans are physically separate from other species and from nonhuman nature; second, that humans are uniquely distinguished by consciousness and agency; and third, that these qualities render humans intrinsically more valuable than other beings [23]. By repudiating these premises, Multispecies Justice expands the scope to encompass not only humans but also animals, plants, ecosystems, and even spiritual or material entities. Animal rights theory has long articulated utilitarian arguments for reducing suffering [105] and ontological claims for recognizing animals as persons or citizens with rights. Biocentric perspectives extend moral consideration to all living beings, including trees and plants [23]. Environmental Justice has linked social equity with ecological sustainability, demonstrating that distributive justice is unattainable without functioning ecosystems [98], while political ecology has emphasized the co-constitution of humans and nonhumans in urban and ecological transformations [23].

The conceptual scaffolding of Multispecies Justice has been elaborated through four interlinked coordinates. It is intersectional, attending to the ways race, gender, class, colonialism, and species intersect to produce vulnerability and marginalization [26, 121]. It is inclusive, grounding autonomy not in liberal individualism but in relationality [83] and in agential realism, which insists that matter and meaning are co-constituted [14]. It foregrounds responsibility and draws on Indigenous epistemologies that emphasize reciprocal responsibilities with rivers, forests, and lands [116] and it is cosmopolitical, expanding political representation beyond the human to include nonhuman actors through concepts, referring to Stengers’ [110], Latour’s [68], and Sheikh’s [104] notions of cosmopolitics. Instead of referring to certain types of justice, such as distributional justice or recognition justice, the Multispecies Justice perspective critiques these categorizations. Their approach is more integrative and relational, emphasizing that justice cannot be adequately understood through frameworks developed for human societies alone. Rather than separating justice into distinct dimensions, they propose a reorientation of justice itself, one that is grounded in the interdependence, vulnerability, and agency of all living beings (human and more-than-human) [121]. However, critics on the concept have argued that justice requires agents, subjects, and institutions and they argue that the hybridism (i.e. everything is entangled) can blur responsibility and it becomes unclear who is morally or politically responsible for injustice [127]. Since the concept is still politically underdetermined, especially when it comes to the question of who speaks, how, and with what authority [24], it is also viewed more as a utopian moral horizon, which is unable to adjucate trade-offs or design institutions [117]. We argue and also are illustrating this with empirical cases, that Multispecies Justice enables us to look beyond the human-centered perspective on justice and this is also it’s core aim.

4.4 Main focus and core questions related to the three justice perspectives

Outlined in the previous sections were the three justice frameworks we use in order to engage with the three stages of AI systems lifecycle. Outlined in Table 1 are some of the fundamental exemplary questions one can use for ethical reflections on cases from those three justice perspectives.

Table 1 Main focus and core questions related to the three justice perpectives
Full size table

5 Discussion: environmental impacts along the AI-lifecycle from different justice perspectives

As we have shown in Sects. 3 and 4 the environmental impacts and resource interdependencies can be analyzed with different ethics perspectives, which help to reveal underlying questions of specific justice claims. In this section we will analyse different concrete cases in the production, operation, and disposal stages along with their corresponding impacts from the three justice perspectives. While recent developments show significant efficiency improvements in data center operations, such as AI-based energy management, advanced cooling systems, and workload optimization, these gains are increasingly offset by a rapid expansion of computing demand. Evidence suggests that despite efficiency gains, overall energy consumption of data centers continues to rise due to cloud expansion, hyperscale infrastructure growth, and AI-intensive workloads [1, 28, 54]. By examining documented cases in countries such as Chile, Mexico, Congo, and Ghana (see Fig. 2), we illustrate how each justice perspective provides a unique lens for observing and revealing different aspects of impacts of AI systems, each from that perspectival level.

Fig. 2
Fig. 2
Full size image

Overview of the empirical cases discussed

All three perspectives are related to different core questions regarding justice, which are summarized in Table 2 and aim to guide researchers in formulating questions and perspectives on ethical reflections on AI systems, and enhance a deeper understanding of the broader impacts of AI systems on the environment.

Table 2 Questions related to AI-Lifecycle Impacts from different justice perspectives and illustrative evidence from cases
Full size table

5.1 Production-related impacts: how mining and extraction exacerbate inequalities

The production stage of AI-system-related technologies involve multitudes of dimensions such as manufacturing, economy, markets etc. But one particular aspect chosen as an illustrative purpose for applying the justice perspectives in this paper is the case of mining for AI system technologies. The reason for this choice is based on the fact that mining industry in general corresponds to 1.7% of the world’s energy use (forecasted to increase to 4–12% in the future, see [10], and to 2–7% of the global GHG emissions [58]. While there are many types of mining, lithium and cobalt mining are among the most prevalent when it comes to AI, which are primarily used in the battery supply chain of AI systems. Two such cases to focus on through the different justice perspectives would be the case of Lithium extraction in Chile’s Salar de Atacama, and Cobalt extraction in Kolwezi in the Democratic Republic of Congo (DRC), both critical suppliers for AI-related production [70, 74].

Due to the rising demand on lithium in the last decade and half (Compare USGS 2012, and USGS 2025), mainly due to green energy transition for batteries but lately also due to its increasing demand for AI applications, there has been a lot of pressure on increasing the extraction of the mineral. Salar de Atacama, originating back to 1962, is one such mine located in Chile and is a part of what is labeled as the ‘lithium triangle (a lithium-rich zone in South America between Atacama in Chile, Hombre Muerto in Argentina, and Uyuni in Bolivia, see [5]), which has been affected by this pressure from multiple aspects. And the other case is Katagana in DRC which is a part of the central African Copperbelt, which is more than 500 km of Lufilian Arc stretching between Kolwezi in DRC and Luanshya in Zambia [82]. This mine is considered as a major source for cobalt mining and located, since more than half of the world’s cobalt production comes from there [13], (Rights and Accountability in Development [RAID], 2024). Therefore, we chose these cases to illustrate how the three justice perspectives mentioned in the previous section can help us in revealing different aspects of these connected problems.

When we look at the Atacama case from an Environmental Justice lens, one perspective to observe is the colonial aspect of the nature of the problem, in which the increase in demand on lithium batteries, particularly in the global north, has two distributional consequences on global south countries. Lithium batteries are used for backup power systems in data centers, edge AI hardware, or sensors. Particularly, the rising demand on lithium batteries not only doesn’t make the locals a primary beneficiary of the products and benefits of the mines compared to the global north, but the rising demand also increases workload and inspection efforts of mining and extraction in those areas [64]. Added to that aspect, there are also concerns over recognitional aspect of these mining practices, as there has been disagreements between the local governmental organizations and the indigenous communities over violation of indigenous rights and for environmental damage [64]. Looking at the Kolwezi case, we can also identify many potential issues. One issue mentioned is the pollution of the water and the air have many negative impacts on both the local population’s health and domestic food yields [82], which raises distributional concerns of age and gender over the benefits and burdens of cobalt extractions in the region. The impact of the pollution is also raising more concerns, as women, newborn and children are more vulnerable to these contaminations than male adults [82, 124, 92].

From a Climate Justice perspective, since lithium extraction is the driving aspect of new batteries, specifically for lowering carbon emissions and better rechargeable batteries, while the end product can be argued that it is better for the environment, the mining procedures leading to that end product with higher rates of extractivism leads to degradation of many aspects of the environment and the ecosystems there in general, thus countering the aimed target it has of a better climatological world [64]. There is the distributional aspect of it as well, since the demand is mostly from the global north, it would also lead mostly to a decrease in the emissions of the global north compared to the global south. As for the cobalt extraction case, there are concerns around the mining practices, as many mention human rights violations in the cobalt mines and advocate for a just transition for the green energy transition [8]. There are multiple concerns over the procedural means of how this mining is done, since there are reports indicating artisanal mining practices and child labour [13, 130].

From a Multispecies Justice perspective, one concerning issue in the case of the Salar de Atacama is that the brines are legally recognized as mining property and the state has dominance on these mines, thus constitutionally denying it as a unique ecosystem or water entity [64]. Not only that, but the over extraction in the mines aren’t allowing natural flow balance of water, since the extraction exceeds annual inflow by 2000 l/s [64]. Another aspect of this is that pressure on lithium extraction led to many other projects to the exploration of other places for mining and extraction, several of them being a part of Los Flamencos National Reserve, which might lead to the endangerment of many animals, specifically flamencos, and lead to a damage in biodiversity [44, 64]. In the case of the Kolwezi in Congo, this damage to the ecosystem has also been reported due to the contamination and pollution of the air and water sources and ecosystems [82].

5.2 Operation related impacts: how the global expansion of data centres fuels local conflicts

The cases of Chile and Mexico described in ‘Where Cloud Meets Cement’ [15] show how the global expansion of data infrastructures reproduces unequal distributions of benefits, power, and recognition. In Chile, the construction of a hyperscale data center in Cerrillos, Santiago, has raised public concern due to its extraction of millions of litres of groundwater in one of the country’s driest regions, where prolonged drought already threatens local livelihoods. In Mexico, the rapid establishment of data centres in Querétaro has intensified pressure on water resources and energy grids, and local residents report shortages and rising costs linked to the enormous cooling demands of these facilities. Companies plan billions of dollars in investments while benefiting from regulatory loopholes that allow them to bypass environmental impact assessments [15].

From an Environmental Justice perspective, the cases exemplify the uneven distribution of environmental burdens and benefits across social and spatial lines [114, 132]. In Mexico the absence of transparent environmental assessment processes further denies affected groups recognition and participation in decision-making, reflecting procedural and recognition-based injustices as defined in Environmental Justice theory [35]. The case highlights how digital expansion, framed as progress, perpetuates colonial and capitalist dynamics that exploit ecologically vulnerable territories in the Global South [52]. In Chile, the proposed Google data centre in Santiago reveals similar inequities, although local resistance and a tribunal ruling partially mitigated them [15]. As shown by Tironi and Albornoz [118], the project’s extensive water demands in a drought-stricken region intensified conflicts over resource allocation [69], making visible how digital infrastructures depend on extractive environmental relations. Public concern over water use during a severe drought led to revisions in the project’s cooling system. Local resistance underscored the role of procedural justice in demanding transparency and participation [35]. Yet, despite this limited victory, water policies still favoured industrial and urban interests, reflecting persistent structural inequalities and the marginalization of affected communities [63]. From an Environmental Justice perspective, the case exposes how regulatory weakness and corporate influence prioritized technological expansion over community and ecological welfare [20, 94].

When it comes to the Climate Justice perspective it becomes clear that the data center expansion amplifies local vulnerability to climate-induced drought and heat stress. While multinational corporations and urban elites benefit from digital growth, rural and low-income communities bear the brunt of resource depletion and ecological degradation. Such dynamics mirror broader global patterns where regions in the Global South absorb the environmental costs of industries that primarily serve wealthier economies [19, 60]. Procedurally, the lack of environmental impact assessments and limited public consultation in Querétaro demonstrate failures in fair participation and representation, central tenets of procedural Climate Justice [50, 135]. Local populations, who experience intensified climate vulnerability, are excluded from governance processes that directly affect their access to essential resources. Corrective or restorative justice remains largely absent, as neither the corporations involved nor the state address the long-term consequences of water overuse or the compounding effects of climate change on already stressed ecosystems [120]. The case exemplifies how technological expansion deepens climate injustice by reinforcing structural inequalities and ecological precarity [98].

From a Multispecies Justice perspective, this development raises some concerns because it emphasizes the interconnectedness of humans and non-human systems, such as water, soil, and ecosystems, and it illustrates that justice must extend beyond benefits for humans [24, 121]. In Querétaro (Mexico), water, aquifers, and ecosystems are treated as resources for human use rather than as entities in a shared system of life. Rather than enabling coexistence and flourishing across species, the current model deepens ecological depletion and social inequity [15]. Through the lens of Multispecies Justice, this case shows a lack of recognition of relational interdependence between humans and non-humans. As Tironi and Albornoz [118] illustrate, in Santiago the data center conflict revealed the need for an ethics of planetarity, recognizing more-than-human relations and territorial justice. Through this lens, community activism made visible the importance of water as a living system, not merely a commodity [121]. Furthermore the loss of biodiversity due to deforestation in ecological sensitive areas, as documented in Rio del Grande, Brazil [16] exacerbates environmental vulnerability that accounts for more-than-human entities. However, the discourse largely remained centered on human water access, leaving the needs of ecosystems and non-human beings implicit rather than explicit.

5.3 Disposal-related impacts: the unequal distribution of electronic waste and its community impacts

Ghana is one of the biggest ports for e-waste disposal [85]. Ghana’s e-waste stream consists mainly of discarded ICT equipment, it also includes significant volumes of used lead-acid batteries (ULABs) from uninterruptible power supply (UPS) systems for servers and backup energy infrastructure [11]. The country’s largest former e-waste hub, Agbogbloshie in Accra, processed an estimated 39% of Ghana’s e-waste and was ranked among the world’s ten most polluted sites due to toxic burning and informal metal recovery [85]. Despite regulation, informal recycling continues to expose workers and nearby residents to lead, mercury and persistent organic pollutants. This reflects a global pattern of externalising the toxic consequences of digital infrastructures to disadvantaged regions. Although the following case is not specific to AI hardware, it illustrates global patterns of hazardous e-waste flows that are likely to intensify as AI accelerates hardware turnover and data centre expansion [119, 131].

From an Environmental Justice perspective, Ghana illustrates a clear case of “toxic colonialism”, a practice in which high-income countries export hazardous waste to lower-income regions [85]. While wealthier economies benefit from digital growth, hazardous e-waste is externalised to countries with weaker regulatory capacity, shifting environmental risks to communities least responsible for producing this waste and resulting in an unequal distribution of environmental burdens and benefits. E-waste enters Ghana through illicit and semi-legal trade channels, often misclassified as “used electronics” to evade environmental regulations, undermining transparency and accountability in global waste governance, which is an evidence of procedural injustice. Moreover, the affected local communities and workers, many of whom are internal climate migrants, have no meaningful participation in decisions over e-waste imports or management [9, 67]. The case demonstrates a recognitional injustice as well. The labour of informal recyclers, who perform essential but hazardous material recovery work, is socially devalued and institutionally stigmatized. These workers are framed as part of the “informal problem” rather than as agents within a globalized resource economy [39]. The Ghana case demonstrates structural injustice with colonial patterns of resource extraction and waste disposal, where the Global South serves as both a site of material extraction and toxic dumping.

Moving to the Climate Justice perspective, the Ghana e-waste case exposes how the ecological consequences of digitalisation intersect with broader patterns of unequal climate vulnerability. E-waste is a driver of climate-related burdens. Informal burning of cables and plastic releases greenhouse gases and acid leaching [38]. These emissions do not originate from Ghanaian consumption but are outsourced from wealthier countries, making Ghana a net receiver of embodied emissions. It demonstrates that in the same way that production burdens may be outsourced, so too may e-waste harms [86], meaning that even in the disposal phase the environmental and climatic costs of digital infrastructures can be relocated to the Global South.

From a Multispecies Justice perspective, the e-waste situation in Ghana extends the site of injustice beyond human communities. The informal recycling of e-waste in Ghana has led to the release of heavy metals and persistent organic pollutants into surrounding soils, sediments and waterways. Analyses of water and sediment in Ghana’s e-waste disposal zones show heavy metal and flame-retardant contamination, indicating that aquatic organisms are at high risk of exposure to these pollutants [57]. Toxic dumping and cable burning at Agbogbloshie have contaminated soils and rivers. The Odaw River, which used to be a fishing site, now is considered to be “dead”, with the loss of aquatic life and vegetation [9]. These impacts show how nonhuman beings are exposed to industrial violence without recognition in policy debates, aligning with Multispecies Justice critiques of anthropocentric environmental governance [24].

6 Conclusion

In this paper, we used three justice perspectives, namely Environmental Justice, Climate Justice, and Multispecies Justice, to reflect on the environmental impacts and their social and ecological consequences by using exemplary contemporary cases in Chile, Mexico, Ghana and Congo. These three perspectives provide complementary lenses for examining the environmental impacts of AI across its lifecycle because each framework captures dimensions of harm and responsibility that the others cannot fully address on their own. Environmental justice foregrounds how environmental burdens and benefits are distributed across human communities, highlighting issues of inequality, exposure, and power in the siting of data centers, mining operations, and waste facilities. Climate justice extends this analysis by situating AI’s energy use and emissions within broader questions of climate responsibility, historical contribution, and intergenerational equity. Multispecies justice further expands the scope by drawing attention to how AI’s extractive and infrastructural demands affect non-human life, ecosystems, and ecological relations that remain invisible in human-centered justice frameworks [97, 113, 30].

Even though we have seen that these impacts are associated with manifold inequalities affecting local communities, vulnerable groups and non-human entities, the fundamental ability of society to act in order to induce change and improve these impacts, thereby reducing their severity, is of utmost importance. There is a wide range of debates within and beyond science on sustainable AI [125] or comprehensive sustainability criteria for AI systems [96], as well as sustainable AI regulation [46]. Building on these insights we want to conclude with approaches to address those impact from three different perspectives, firstly within the realm of model development & training, second related to the infrastructural prerequisites and third more broad policy approaches to address those issues. Regarding Model Development and Training the Sustainable Machine Learning Design Pattern Matrix [71] is a valuable approach, which shows that sustainability in AI cannot be achieved by optimizing training efficiency alone, but is strongly related to early-stage design choices, such as problem formulation, data selection, and model sufficiency, rather than late-stage technical optimizations alone. Data-centric practices such as limiting data volume, reducing dimensionality, and avoiding unnecessary retraining can significantly lower energy consumption without sacrificing societal value. By emphasizing sufficiency-oriented performance targets and efficiency-aware model selection, design choices should discourage unnecessary model complexity and reduce environmental impacts when additional performance gains provide only marginal value [71]. This resonates with approaches on Green AI [100], which emphasize smaller, task-appropriate models or selecting appropriately sized models that maintain sufficient utility while consuming significantly less energy than oversized architectures [129]. When it comes to the infrastructural prerequisites namely hyperscale data centers and related network infrastructures, energy use can be reduced through mix of passive strategies (site/climate choice, building airflow design, “free cooling” where outside conditions allow) and active strategies (high-efficiency chillers, optimized airflow containment, and control/automation) to lower cooling demand and overall Power Usage effectiveness [21]. But the impacts that arise from the infrastructural prerequisites can not only be solved by design but also require forms of governance that extends beyond algorithmic performance to address the socio-technical-ecological systems that AI is embedded in [102].

In conclusion, effective AI governance must make environmental costs visible through transparency and reporting of energy, carbon, water, and hardware use [32, 112] and there is need for a shift in the regulatory focus, which integrates sustainability requirements, such as environmental impact assessments, consumption limits, and sustainability-by-design principles [45]. Hyperscale data centers should be sited in locations that are not exposed to water stress, drought, flooding, or ecological vulnerability, and that can support cooling and operation without placing additional strain on local communities or ecosystems. At the same time, the infrastructures that sustain AI, hyperscale data centers, electricity grids, water systems, mineral supply chains, and waste streams, must be governed as integral parts of AI systems rather than treated as externalities [31].

Future research should further develop the concept of AI as a socio-technical-ecological system by theorizing how algorithms, infrastructures, and ecosystems are co-constitutive rather than separated. In particular, empirical research is needed to trace how AI infrastructures produce unequal environmental burdens for communities in different spatial and social contexts.