I would like to thank my wonderful colleague Nancy Piñeiro for several years of invaluable discussions of the role of translation in supporting communities engaged in socioenvironmental conflicts. Her work in translating and co-producing knowledge in support of community resistance to fracking in Argentine Patagonia has been a constant source of inspiration and insight. In August 2024, I had the privilege of working with Nancy on a joint conference presentation in Buenos Aires entitled Traducción y crisis climática: desafíos y solidaridades Norte-Sur [Translation and the climate crisis: Challenges and North-South solidarities]. Our conversations and joint work have greatly influenced and informed this article.

I would also like to thank the anonymous reviewers for their exceptionally thorough and incisive feedback on the originally submitted version of this paper, as well as this issue’s co-editors, Sue-Ann Harding and Robert Neather, who have also provided excellent guidance and been an absolute joy to work with. Moreover, I extend my gratitude to all other colleagues whose interest in and engagement with my work have helped refine my ideas. I am solely responsible for any errors or perceived shortcomings in this paper.

I would like to commend the vital efforts of those activists who are already engaged in relevant struggles, many of whom are mentioned in the paper. I continue to draw inspiration from the work of groups such as AI + Planetary Justice Alliance, Amazon Employees for Climate Justice, Arte es Ética, Guerilla Media Collective, Tu Nube Seca Mi Río, and many others.

AI technologies’ substantial ecological costs are eliciting increasing concern in academic research and international media. Though few would argue against emissions reductions wherever possible, others may be sceptical of the extent of these technologies’ impact on the climate crisis. After all, the entire information and communications technology (ICT) sector accounts for just 2-6% of global emissions, with AI representing a smaller fraction (Luccioni & Hernandez-Garcia, 2023, p. 1). Nevertheless, recent trends indicate that AI may be emerging as an increasingly integral part of the global economy. Following ChatGPT’s release in late 2022, large language models (LLMs) in particular have become a prominent fixture in the AI landscape and across the wider economy. Automated language translation—most often facilitated by neural machine translation (NMT) but now also performed via LLMs—is also frequently upheld as a premiere example of the triumphs of AI (see Srnicek, 2022, pp. 242–243; Wu et al., 2022, p. 5). Amid the worsening climate crisis, these conditions have prompted language professionals and translation scholars to consider the environmental dimensions of the field’s technological practices.

NMT was just beginning to enjoy widespread adoption when Cronin (2017) coined the term eco-translation—a concept that encompasses “all forms of translation thinking and practice that knowingly engage with the challenges of human-induced environmental change” (p. 2). Even then, NMT’s steep energy requirements were already cause for concern, with Cronin (2017) advocating for a return to “low-tech translation” (p. 106). Needless to say, this future has not materialized. Post-editing of AI-generated text is now a widespread practice in the language services industry, and translators are generally pressured to accept these lower-paying jobs due to diminished bargaining power in the industry (do Carmo, 2020; Moorkens, 2020; Carreira, 2023). These practical constraints hint at the seeming impossibility of a potential embrace of “low-tech translation” rejecting AI’s ecological costs. Reconciling environmental concerns with technological benefits, Moorkens et al. (2024) propose a sustainability model for translation automation that balances impact on people, planet, and performance. Still, the authors take care to emphasize the extreme difficulty of untangling the interplay between these technologies and wider socioeconomic structures (Moorkens et al., 2024, p. 10). Within or beyond the translation field, reckoning with AI’s environmental harms necessarily entails contending with the powerful structural forces that shape these technologies’ use.

Macro-level patterns in consumption and technology use rarely acknowledge the socioenvironmental conflicts over extractive, production-oriented projects and the stark power asymmetries that define them. Against this background, Piñeiro (2022) promotes “counterhegemonic translation” (la traducción contrahegemónica) as a way to support collective efforts to challenge powerful actors perpetuating socioenvironmental harms. This approach gives central importance to translation as a process through which those “on the frontline of the fight against extractivism” (en la primera línea de la lucha contra el extractivismo) are able to “access the information they need, secure key materials, and build networks of international solidarity” (acceden a la información que necesitan, se apropian de materiales claves y construyen redes de solidaridad internacional) (Piñeiro, 2022, p. 159, my translation). Piñeiro’s (2026, this issue) own translation practice and ethnographic research involving community resistance to fracking in Argentine Patagonia provides a quintessential case study for counterhegemonic translation. Similarly aspiring to merge resistance with research practices, Boéri and Baker (2025) urge defiance against the extractive nature of AI translation, a tool whose commercial iterations proffer widespread language access while quietly extracting data and value from users. Adopting this ethos of resistance, this article develops a framework for eco-translation practice that confronts the extractivist forces perpetuating AI’s environmental harms both within and beyond the translation field.

The following section reviews recent literature to underscore the multifaceted and far-reaching nature of AI’s environmental impacts, whose complexity and severity are still not widely grasped in the translation field. Subsequently, the paper presents three distinct narrative framings of proposed approaches to solving this issue, evaluating the suitability of each as a basis for eco-translation practice in this area. The next section then maps the eco-social relations along AI’s global production chain, drawing upon Princen et al.’s (2002) concept of distancing to indicate how resource consumption along this global production chain is insulated from public objections. Finally, the paper proposes actions for reducing distancing along this chain and thereby resisting AI’s environmental harms both within and beyond the translation profession. To this end, the article differentiates between translators’ potential vocational resistance and structural resistance to AI’s environmental harms—the former oriented towards translation as a profession, and the latter oriented more broadly towards translation as a practice.

Understanding the scope of AI’s ecological impacts

While the problem of AI’s ecological costs is certainly gaining traction, translation professionals and scholars may still not fully grasp the nature and scale of these harms, as research is often diffuse and/or overly technical. This section surveys recent works to provide an accessible overview of the main ecological tolls linked to AI technologies.

Carbon emissions

AI models’ carbon emissions are commonly attributed to their processes’ massive energy requirements, which, depending on the location in which these processes are carried out, often rely on electricity generated from fossil fuels. Different regions’ energy grids have vastly different energy mixes—i.e., combinations of renewable and/or non-renewable sources drawn upon to meet energy demands (Shterionov & Vanmassenhove, 2023).

AI’s direct energy consumption and its subsequent emissions may be divided into two primary workload phases: training and inference. Models’ initial training processes are extremely computationally intensive and therefore highly carbon-emitting. Energy requirements for training have risen in tandem with model sizes, as increasing model size has generally been the primary means of improving performance. Steadily increasing energy requirements have led to growing emissions, though the multiplicity of other influencing factors prevents a perfect correlation between the two (Varoquaux et al., 2024, p. 6). While model training generates substantial emissions, cumulative emissions from inferencing can be even greater.

Inferencing refers to a model being used for its intended purpose, such as when an NMT model translates a source-language input or when an LLM generates an output. A single inference consumes a fairly small amount of energy, thus generating few emissions. But while model training is typically a one-off event, inferencing is ongoing: the aggregate energy consumption of recurrent inferencing can easily surpass the energy consumed during training, particularly for widely used models (Luccioni et al., 2023). When comparing different models on the same tasks, the general-purpose models are higher-emitting than task-specific models, and “decoder-only models [e.g., LLMs] are slightly more energy- and carbon-intensive than sequence-to-sequence models [e.g., NMT] for models of a similar size” (Luccioni et al., 2023, p. 13). These differences in energy consumption are an important takeaway, given the AI industry’s growing preference for general-purpose models (Luccioni et al., 2023). General-purpose AI models are also closely tied with the rapid construction of new data centres, whose emissions pose significant health problems for nearby communities (Han et al., 2024). Nevertheless, carbon emitted during the training and inference phases only reflects operational emissions. AI’s embodied emissions are also critical to consider.

Embodied emissions represent carbon emitted during any upstream and downstream activities in a product’s supply chain, including materials extraction, production, transport, manufacturing, and hardware disposal (Luccioni et al., 2023, p. 3). These emissions are important to consider when assessing technologies’ carbon impacts. For instance, steel and cement are essential to decarbonizing infrastructures, yet the production of these materials is itself highly carbon-intensive (Buller, 2022, p. 66). Additionally, the vast majority of copper mined worldwide goes towards electronics manufacturing, with copper production accounting for over 50 million tons of CO₂-equivalent emissions per year (Falk et al., 2024, p. 5).

Developers frequently encounter a tension between AI’s operational and embodied emissions. Reductions in operational emissions often result in larger or comparable increases in embodied emissions, as components of state-of-the-art hardware usually contain far more embodied carbon (Gupta et al., 2022; Wu et al., 2022). Accordingly, Wu et al. (2022) predict that embodied emissions will come to represent the majority of AI models’ emissions, while Gupta et al. (2022) anticipate this trend spreading across the entire ICT sector. A recent study from Google refutes these conjectures. The Google researchers apply a “cradle to grave” lifecycle analysis of successive generations of the company’s proprietary AI hardware, claiming that operational emissions far surpass embodied emissions (Schneider et al., 2025, pp. 8–9). As detailed below, however, far greater transparency is needed to verify these claims.

Water consumption

AI’s associated water consumption is also causing serious concern. Water consumption refers to water being used in such a way that, although it may eventually still integrate back into water cycles, it becomes unavailable for immediate use, such as through evaporation or ingestion by humans. The overconsumption of water intensifies water cycles and in many of the world’s regions exacerbates water scarcity—one of the planet’s most pressing challenges (United Nations, 2024).

AI models’ water consumption may be separated into direct/onsite (i.e., in data centres) consumption and indirect/offsite consumption associated with energy production (Siddik et al., 2021; Li et al., 2023). Onsite water consumption is necessary for cooling data servers to prevent overheating due to the exorbitant computational intensity. As the most common cooling method, cooling towers draw water to release heat via evaporation (Li et al., 2023, p. 4). Like carbon emissions, direct water consumption is heavily dependent on timing and location, in addition to many other variable factors involving water cycles and airflow (Li et al., 2023, p. 5).

Though data centres’ aggregate water consumption may appear relatively small (Kwet, 2024, p. 116), Big Tech companies continue to operate and construct numerous data centres in some of the world’s most water-scarce regions. A recent investigative report by The Guardian and SourceMaterial revealed that Amazon, Google, and Microsoft have 38 active data centres and 24 planned data centres in water-scarce areas around the world (Barratt & Gambarini, 2025). In Latin America, many data centres are clustered around metropolitan areas in Chile and Brazil—two countries struggling with severe water shortages. The proliferation of AI models is significantly increasing demands on data centre processes, thus directly increasing onsite water consumption and exacerbating these issues.

Offsite water consumption mainly refers to water consumed during energy production. The level of water consumption required for electricity generation varies greatly according to the energy source and technological processes applied, but it is generally high (Siddik et al., 2021). For instance, Li et al. (2023, p. 3) estimate that training GPT-3 in the US consumes 2.8 million litres of water from electricity generation alone. Energy production represents another key component in global water scarcity, as it accounts for between 10 and 15% of water withdrawn globally (United Nations, 2024, p. 4).

Pollution from mining rare minerals

Data centres, user devices, and renewable energy infrastructures all require rare minerals and other materials whose extraction causes serious harms. Ecosystems are often degraded by mining elements needed for microchips, including copper, gallium, germanium, indium, lithium, tantalum, and tellurium (Falk et al., 2024, p. 6). Mining these materials creates hazardous waste that often harms communities by contaminating local sources of food and water (Carayannis et al., 2025).

Among the most essential rare minerals are lithium and cobalt, whose extraction is deeply implicated in serious cases of environmental degradation, labour exploitation, and human rights violations in the Global South. Much of the world’s lithium supply is mined in the so-called Lithium Triangle, spanning Argentina, Bolivia, and Chile. As a highly water-intensive process, lithium mining in the Atacama Desert in northern Chile has drastically intensified water scarcity, polluted ecosystems, and damaged biodiversity in the region (Lehuedé, 2024). The Democratic Republic of the Congo supplies the overwhelming majority of the world’s cobalt, essential in lithium-ion batteries. The country’s cobalt mining industry is notorious for rampant labour and human rights violations, including forced child labour and numerous fatalities due to dangerous working conditions (Brodkin, 2024).

Although it may be difficult to dispute the seriousness of the environmental harms outlined in this section, it is far less certain how a solution to these complex issues should be formulated. Even more ambiguous is the role that translators—professional or otherwise—might play in addressing these environmental harms. In order to provide clarity on these points, it is necessary to examine how competing framings of the problem imply solutions of varying efficacy.

Framing the problem of AI’s ecological harms

As Baker (2019) argues, narrative framings are indispensable to conflict, and any form of translation oriented towards social change must identify and challenge the dominant narrative sustaining the status quo. Amid the increasing attention paid to AI’s environmental tolls, there are several competing narratives proffering radically different solutions. This section scrutinizes three overarching narratives, highlighting the efficacy of their proposed solutions and their suitability as a basis for eco-translation practice in this area.

The false promises of the corporate sustainability paradigm

Many developers and cloud providers embrace a controversial practice called carbon offsetting. Carbon offsetting schemes are transactions in which polluters supposedly counterbalance their emissions by investing in sustainability initiatives alleged to prevent the equivalent amount of firms’ attributed emissions from entering the atmosphere, thereby achieving “net zero” emissions. In this view, carbon offsets are rendered commensurate with firms’ actual emissions via standardized units, then commodified and traded on carbon markets. Historically, the world’s most powerful technology corporations have embraced carbon offsets as their primary approach to sustainability (Patterson et al., 2022, p. 24), yet the supposed one-to-one countervailing effect between actual emissions and firms’ offset purchases is highly questionable.

Carbon offsets’ alleged emissions-negating effects are rooted in the generation of some quantity of renewable energy to fulfil energy needs that would have supposedly otherwise drawn upon non-renewables. Offsetting neither decreases atmospheric carbon in absolute terms nor reduces emissions generated by firms. Purchased offsets merely pre-empt hypothetical future emissions that most often would have occurred on sites and in timescales that are completely different from firms’ actual carbon-emitting operations. As critics note, it is impossible to verify or accurately quantify these negated emissions, since they reflect hypothetical scenarios reflecting the whims of an unregulated, largely opaque carbon market; numerous independent studies have found massive flaws in prominent offsetting programmes’ methods and claims (see Temple, 2024). Similarly, corporate pledges to become “water positive”—such as those of Microsoft and Google—depend on commercially-convenient projects to replenish selected water basins and simplistic accounting based on false equivalencies, much like carbon offsetting (World Wildlife Fund, 2022). These pledges also neglect off-site water consumption, which often constitutes a significant proportion of AI’s overall water consumption (Falk et al., 2024, p. 5).

Recognized as a world leader in sustainability innovations, Google has also declared its intention to achieve “24/7 carbon-free energy” (CFE) for all office and data centre operations in light of the unreliability of carbon offsets (Bryan et al., 2024). However, this scheme is misleading, as it does not mean that all operations will run on renewable energy at all times. Instead, the company will simply match its emissions by concurrently purchasing renewable energy credits for the same energy grid in the same hour (Bryan et al., 2024). Moreover, as a report by the environmental organization STAND.earth (2023, p. 20) observes, Google seemingly has no plans to decarbonize its supply chain, even though these embodied emissions constitute a significant portion of its total emissions. As indicated above, though, a more recent report by Google researchers asserts that the company’s operational emissions far outweigh its embodied emissions (Schneider et al., 2025). Still, Google’s estimates rely on primary data on supply chain emissions that remain inaccessible to third parties, and the aforementioned CFE scheme factors heavily into these calculations.

Independent investigations and critical reporting make evident the inherent conflict of interest in tech companies’ voluntary environmental disclosures. Big Tech companies have dedicated substantial resources toward producing self-affirming research and lobbying the Greenhouse Gas Protocol—which oversees carbon accounting methods—to adopt misleading accounting and disclosure protocols (Bryan et al., 2024). One independent investigation estimates data centre emissions to be 662% higher than reported by Big Tech cloud providers (O’Brien, 2024). Water consumption may also be drastically underestimated in voluntary disclosures; the International Energy Agency estimates data centres’ global water consumption in 2023 to be 560 billion litres—much higher than originally thought—and anticipates this figure to increase to 1.2 trillion litres by 2030 if the current trajectory holds (IEA, 2025, p. 242).

The limits of disclosure and technological efficiency

Aside from these corporate sustainability practices, researchers and industry professionals have promoted public awareness and technological efficiency as primary strategies for mitigating AI’s environmental costs. Several online tools have been developed for estimating models’ carbon emissions, but these emissions calculators are found to vary widely in their estimates, and they generally underestimate models’ true emissions by discounting embodied emissions (Bannour et al., 2021). Naturally, emissions calculators rely heavily on guesswork due to the general lack of transparency regarding digital and energy infrastructures (Luccioni & Hernandez-Garcia, 2023). While leaderboards and similar efforts are perhaps encouraging, it is particularly dangerous to “assum[e] that disclosure in and of itself will spur the markets to drive a sustainable future” as there is no evidence to support this notion (Buller, 2022, p. 167). The World Wildlife Fund (2022) similarly warns against equating disclosure with meaningful action, and stresses that consumption and replenishment metrics cannot account for the highly delicate qualities of water that influence local biodiversity.

Technology efficiency may present an enticing approach to AI’s environmental costs. But the variability of earth systems can also counteract efficiency in any one particular dimension. For instance, choosing optimal times and locations for reducing models’ carbon emissions may dramatically impair water efficiency, as data centres operating in sunny areas abundant in renewable solar energy often require more water for cooling purposes due to higher temperatures (Li et al., 2023, p. 7). The efficiency angle also neglects the much more pressing and complex reality of AI’s aggregate ecological costs in the face of these technologies’ unimpeded expansion. This disconnect relates to a phenomenon commonly referred to as the Jevons paradox or the rebound effect: technologically driven efficiency gains in resource use may counterintuitively result in greater aggregate resource consumption, since increased efficiency raises demand. The AI landscape is believed to exhibit this phenomenon, as exploding user demand for this emerging general-purpose technology seems to eclipse improvements in model efficiency (Varoquaux et al., 2024; Bhardwaj et al., 2025 ; Luccioni, Strubell, & Crawford, 2025).

Heightened demand for cloud services has spurred the rapid construction of new data centres and propelled energy consumption upward. The total number of hyperscale data centres worldwide has doubled over the past five years, and the combined capacity of facilities is expected to increase nearly threefold over the next six years (Thorne, 2024, pp. 13–14). As a result of these developments, Big Tech’s energy consumption and emissions totals are skyrocketing: Microsoft’s carbon emissions are up 30% compared to its 2020 baseline, while Google’s have increased 48% compared to its 2019 baseline; these figures cast serious doubt on the companies’ near-term climate goals (Bryan et al., 2024).

AI and planetary boundaries

Counter to these techno-solutionist approaches, a more critical framing of AI’s planetary harms evaluates the cumulative impacts of these technologies on Earth’s delicate systems and resource limitations, drawing upon the planetary boundaries framework (see Falk et al., 2024; Bhardwaj et al., 2025). In their seminal paper, Rockström et al. (2009) introduced the concept of planetary boundaries—quantitative estimates of indicative thresholds, such as atmospheric carbon concentration and global freshwater use, that humanity must maintain in order to avoid irreversible changes to planetary processes and mass ecological catastrophe. The concept has since been widely accepted in scientific research and made its way into mainstream discussions of climate change.

At the same time, the planetary boundaries framework has seen key contributions from the social sciences. Brand et al. (2021) insist that social scientists’ perspectives are “essential for going beyond the diagnosis of the transgression of planetary boundaries” (p. 265, original emphasis) to reveal the underlying social forces driving humanity’s collective breach of these thresholds. They cite a litany of evidence attributing the push above planetary boundaries to capitalism’s central growth imperative, observing that powerful corporations’ compulsion to grow at all costs leads them to “actively shape and condition demand for their output” as well as externalize ecological costs and evade regulation (Brand et al., 2021, pp. 270–271).

Recently, critical research has weighed the impacts of AI on planetary boundaries, highlighting the social dynamics driving this problem (Falk et al., 2024; Bhardwaj et al., 2025). Others observe that the tech industry’s concentrated economic power is a massive factor in shaping global consumption. Kwet (2024) goes so far as to assert that tech giants’ economic dominance is directly responsible for “radically violat[ing] planetary boundaries” by serving as “[t]he central driver of limitless growth on a finite planet” (p. 223). Hogan and Blue (2024) argue that, by way of the corporate sustainability paradigm outlined above, “big tech (increasingly as big cloud) companies position themselves as the best custodians of nature and of natural resources” (p. 33) by reinforcing their unilateral decision-making power over land and resource use. The concentration of power among these tech giants presents an unambiguous hegemony to confront in formulating a counterhegemonic translation agenda centred on AI’s environmental costs. In order to postulate a specific set of resistance strategies for translators, though, a more in-depth look at the nature of the power relations governing the AI industry’s resource consumption is needed.

Mapping the eco-social relations underpinning AI

Material and energy throughputs threatening planetary boundaries are most commonly expressed in terms of production and consumption—the complementary pillars of global economic activity. However, it is not readily apparent what is “produced” or “consumed” across the AI field, or what factors shape its constituent forms of resource consumption. Srnicek (2022) provides a helpful typology. AI providers—who may perhaps be considered as the primary “producers” of AI—are the tech giants that control “the ownership and provision of the means of production of AI” and thus support other firms’ AI models (Srnicek, 2022, p. 243). This group is largely synonymous with Big Tech, denoting cloud computing giants such as Amazon, Google, and Microsoft. AI consumers are “those who purchase and use AI services from others” and face the “competitive pressures driv[ing] the adoption of machine learning across the economy” (Srnicek, 2022, p. 244). Still, these broad distinctions do little to clarify responsibility for the environmental impacts outlined in Section 2. The relation between translators’ potential “consumption” (i.e., use) of AI and the various forms of material resource consumption comprising AI’s cumulative ecological impacts requires further examination.

Producing and consuming AI

Consumption may be broadly conceptualized as any form of “human activity that ‘uses up’ material, energy, and other valued things” (Princen, 2002a, p. 29). The emphasis on disclosure and efficiency described earlier suggests that an ad hoc form of green consumerism has taken shape in the AI field. Green consumerism rejects the need for deeper structural changes to the political economy; instead, it supposes that sufficiently-informed individual consumers—e.g., AI consumers such as translators—will induce more environmentally-friendly product offerings and manufacturing practices from producers via some critical mass of uncoordinated market choices or consumption habits (Akenji, 2014). Here, end consumers are treated as the ultimate determiners of resource extraction, exercising their authority by making implied demands of producers via their consumption habits. However, green consumerism disregards the multiplicity of complex factors influencing production and consumption patterns, turning isolated consumers into “scapegoats” for a society-wide problem, as argued by Akenji (2014).

To the contrary, Princen et al. (2002) contest that consumer choices are “subject to structural features that often make it convenient, rewarding, and even necessary, to increase consumption” (pp. 14–15). It is therefore crucial to situate consumers’ decisions within a “larger web of social relations” to critically examine “the influences on consumption choices, including the location of power in structuring those choices” (Princen et al., 2002, p. 15, emphasis added). Rather than assessing individual actions in terms of resource efficiency, the task then becomes determining the manner in which power shapes society’s interdependent forms of production and consumption (Brand et al., 2021). This perspective enables connections between “socially embedded consumers” positioned within “linked chains of resource-use” such that all forms of consumption along these chains are considered, instead of merely the “downstream node” of end consumers’ demand for a given product (Princen et al., 2002, p. 16). It is therefore imperative to scrutinize the interconnected nodes constituting “the supply chain of AI” as a “complex, global and opaque mechanism” that distorts public perception and accountability for environmental harms (Valdivia, 2024, p. 2).

Ecological stewardship dissipates as consumers become more detached from the consequences of their decisions. Princen et al. (2002) refer to this concept as distancing, which denotes “the increasingly isolated character of consumption choices as decision makers at individual nodes are cut off from a contextualized understanding of the ramifications of their choices, both upstream and downstream” (p. 16). This isolation mutes “negative feedback loops” by severing the “information flow that connects the dynamics of ecosystem support to resource availability”; isolated consumers along the production chain thus develop the perception that resources are “infinite or infinitely substitutable” (Princen, 2002b, p. 128). As detailed below, there is an extremely high degree of distancing along AI’s global production chain.

There are four interrelated dimensions of distancing—geography, culture, (bargaining) power, and agency (Princen et al., 2002). Often in concert with geographic distance, cultural distance “block[s] ecological feedback by inhibiting information flow from extraction to consumption decisions” (Princen, 2002b, p. 119). These paired dimensions establish the concept’s relevance to translation: language frequently provides a barrier to the critical information flows necessary for organizing local resistance to extractive projects, making translation an indispensable component of socioenvironmental conflicts (Piñeiro, 2022). Bargaining power denotes market power imbalances in which resource consumption is dictated by a limited number of disproportionately powerful firms (i.e., monopolies or monopsonies). The agency dimension refers to the number of intermediaries between primary materials and the end-point consumer, which can dilute the perceived responsibility for society-wide consumption patterns. Together, these forms of distancing ensure that environmental degradation along production chains remains impervious to wider social forces that would reject these costs. Princen (2002b, p. 129) argues that lowering distance along any of these four dimensions will reduce the dispersion of accountability among production chains’ webs of decision-makers, thereby encouraging ecological feedback loops that inhibit the overconsumption of resources and prevent environmental degradation.

This section has argued that translation may assist in decreasing distance across AI’s global production chain. As explored in the following section, the high degree of distancing in the AI industry offers many such opportunities.

Distancing along the global production chain of AI

For the purposes of this article, it is helpful to analyse distancing within AI’s global production chain by working backwards from the end node of post-editing AI translation output. The portrayal of translators’ positioning in this chain enables the differentiation of strategies to minimize AI’s environmental costs within the translation industry (referred to as “vocational resistance”) from strategies to challenge the broader structural forces escalating AI’s aggregate harms in society (referred to as “structural resistance”).

Professional translation is largely mediated by language service providers (LSPs)—firms that receive and manage translation requests from clients, then outsource actual translation work to a pre-approved pool of freelance translators. This trend has been commonplace in the translation industry for quite some time now, with more and more of translation projects’ constituent tasks being divided and (semi-)automated (Moorkens, 2020). Alongside macroeconomic conditions such as market consolidation (Carreira, 2023) and increasing platformization (Fırat et al., 2024), the prevalence of NMT in translation workflows has therefore applied a strong downward pressure on translator pay. LSPs are incentivized to increase profits by decreasing labour costs (i.e., translator pay), which, in turn, pressures translators to recuperate these losses by increasing their productivity in post-editing AI output (see do Carmo, 2020). Opportunities to counteract these developments via collective bargaining are lacking (Fırat et al., 2024). In this manner, the translation industry reflects a high level of distancing: translators have little control over the conditions of their labour, let alone the conditions under which the AI models they engage with are created and deployed.

LSPs have increasingly turned to AI-powered services—traditionally anchored by NMT, but now also integrating LLMs—and dictated the terms of translators’ engagement with these new technologies (Moorkens et al., 2024). Notably, the collection of language data for training and customizing AI models has proven to be an enormously lucrative business venture since the popularization of NMT (Diño, 2018). This innovative revenue stream depends on the conversion and commodification of translators’ finished work into AI training data, relying on the absence of any contractual obligations (e.g., copyrights) to compensate or concede ownership rights to the data’s original producers (Moorkens & Lewis, 2019). The appropriation of translators’ work as training data for AI models exemplifies the asymmetry of bargaining power between translators and LSPs (Boéri & Baker, 2025). Still, while data has been a key focus in AI development, it is hardly the only input necessary for these technologies.

What is perhaps even more essential is access to compute—the industry’s term for the material infrastructure (primarily hardware, but also software) needed to handle AI workloads. Most AI models operate via third-party cloud computing facilities due to the exorbitant costs of locally processing AI workloads, since they may be more cheaply and efficiently outsourced to major cloud providers—Amazon’s AWS, Google Cloud, and Microsoft Azure. Cloud computing facilities represent a specific type of data centre providing remote, on-demand computing services that enable intensive digital operations—e.g., compute and data storage—to be conducted offsite. The aforementioned trio of tech giants currently controls the overwhelming majority of the global market for cloud computing (Thorne, 2024).

As firms in diverse sectors scramble to take advantage of the perceived profitability of AI, market competition intensifies the already-strong reliance on cloud providers across the economy, with these corporations serving as gatekeepers to the compute needed for AI (Srnicek, 2022). These providers offer some opportunities for LSPs and other AI-dependent firms to mitigate AI’s ecological costs. For instance, Google Cloud offers users the option to choose data centre locations for their AI workloads based on the availability of carbon-free energy1 as well as a more comprehensive tool for estimating their overall carbon footprint.2 However, as indicated in previous sections, these features paint a highly incomplete picture of AI’s ecological tolls and do nothing to curb or dissuade growing aggregate AI use. In fact, Big Tech’s massive investments in AI infrastructures are expressly counting on an enormous, society-wide uptake of AI use; many tech firms are imposing AI-powered features onto widely used digital services and apps that comprise everyday users’ online experience, simultaneously concealing or disregarding the compounded environmental costs of these manipulated digital design practices (Beignon et al., 2025). The dynamic between AI providers and LSPs—or other “AI start-ups” (Srnicek, 2022, p. 243)—therefore reflects a high degree of distancing: market competition encourages LSPs to increase their AI use, while AI providers’ market dominance ensures that these firms have little insight or recourse to influence AI’s ecological costs, short of abstaining from these technologies altogether. In this manner, there is a blatant tension between the commercial incentives driving AI’s widespread adoption and any perceived responsibility to moderate or disclose its associated ecological tolls, making the data centre/cloud computing industry a key target for counterhegemonic translation practices. With data centres serving as the bedrock of AI infrastructures and the concentration of power among cloud providers, this particular node holds enormous potential as a frontline for resisting AI’s ecological harms.

Data centres are already emerging as a principal site for contesting resource consumption in AI’s highly distanced global production chain (Lehuedé, 2022; Jansen & Havana, 2024; Barakat et al., 2025). A report from The Maybe (Barakat et al., 2025) offers a panoply of case studies showcasing local communities’ resistance to the construction of data centres around the world, including sites in Chile, the United States, the Netherlands, Mexico, and South Africa. Although each community is shown to organize resistance around a highly localized and unique set of socioenvironmental issues, two major challenges persist across all five case studies: 1) public perceptions of data centres are unilaterally shaped by companies and governments, and 2) there is frequently little to no transparency regarding planned data centres and their projected impacts on surrounding communities (Barakat et al., 2025, p. 39)3. The accelerated build-out of data centres is the most influential force in AI’s global production chain, propelled by Big Tech firms’ extraordinary accumulation of capital.

This capital accumulation affords Big Tech considerable power over the extraction of fossil fuels and materials needed for data centres’ operations. In countries with favourable business environments such as Ireland, AI providers aggressively pursue public-private partnerships that embed their operations within public energy infrastructures and make states more dependent on their services (Brodie, 2020). Across the United States, natural gas serves much of data centres’ increasing electricity demand for AI (IEA, 2025, p. 75). Additionally, AI providers directly support fossil fuel extraction (Hao, 2024), and Big Tech’s financial investments are closely intertwined with fossil fuel companies (Buller, 2022, p. 130). These realities complicate the image of Big Tech as global leaders in sustainability. The power disparity between Big Tech and other actors in the AI production chain represents extreme distancing, most prominently displayed and maintained in tech giants’ dominance of the global data centre market.

The manufacturing of graphics processing units (GPUs) constitutes another crucial node, just upstream and in close proximity to the data centre industry. Dominated by NVIDIA, the enormously profitable US chip manufacturer, the GPU industry consumes large amounts of silicon, copper, and many other materials extracted from across the world (Valdivia, 2024). The capacity and responsibility for major cloud providers to eliminate socioenvironmental harms in these materials’ supply chains has been a major point of contention. As discussed previously, the cobalt mines in the DRC are rife with human rights violations. A 2019 lawsuit alleged that Apple, Alphabet (Google), Dell, Microsoft, and Tesla control some 80–85% of the DRC’s cobalt supply chain, and thus are culpable in the cobalt mining industry’s inhumane treatment of workers; however, US courts dismissed the case in March 2024, ruling that the companies should not be held legally responsible (Brodkin, 2024). The plaintiffs’ legal counsel argued, paradoxically, that the tech firms claim to have explicit “zero tolerance” policies for child labour in their supply chains while also claiming that they have no ability or obligation to monitor this issue (International Rights Advocates, 2024). As major purchasers of these primary commodities (mediated through firms dealing directly with miners), Big Tech undoubtedly plays a major role in generating global demand for lithium, cobalt, and the many other rare minerals needed for digital infrastructures, despite their apparent lack of legal responsibility to shape these forms of resource extraction. The opacity of the socioenvironmental impacts of the extraction of data centres’ composite materials makes it nearly impossible for actors at later nodes in AI’s global production chain to appraise the full ecological ramifications of these technologies. Sites of primary materials extraction in the AI production chain therefore also reflect high distancing.

Finally, the oft-overlooked terminal node of the AI production chain is the disposal of e-waste, as material-intensive data centre hardware often becomes defunct or obsolete after several years. The critical minerals and other materials comprising this decommissioned equipment generally make for toxic waste, bringing adverse effects to the local environments in which it is disposed—whether properly or improperly. With GPU manufacturers continuously working to improve these vital and highly lucrative products, GPUs are particularly prone to rapid obsolescence, often shipped to Global South countries such as Kenya or Ghana and also to China to be incinerated (Carayannis et al., 2025, p. 10). Amid the ongoing wave of AI expansion, the worldwide demand for e-waste disposal is growing rapidly, though it remains a largely informal market. The lack of formalized employment in e-waste disposal forestalls much-needed worker protections and standardized disposal procedures, resulting in the routine mishandling of this toxic waste (Falk et al., 2024, p. 6). Though global e-waste flows are not tracked systematically, it appears that Africa, South America, and southeast Asia are the primary recipients (Falk et al., 2024, p. 8). These conditions perhaps best exemplify the stark geographic disparities inherent in the AI production chain, with benefits of AI primarily accruing to Global North corporations while hidden environmental and public health costs are borne by Global South workers and communities. Such distancing epitomizes the dilution of responsibility for consequences stemming from the interlinked forms of resource consumption underpinning AI technologies.

As shown in this section, the highly distanced forms of resource consumption along AI’s global production chain result in an apparent diffusion of responsibility for its cumulative ecological harms. Resource consumption at each node is heavily influenced by socioeconomic factors (e.g., market competition). To redress AI’s environmental tolls, it is necessary to reach beyond the limited contexts in which end users such as translation professionals encounter these technologies, instead confronting the structural forces driving AI’s increasing use across society.

It is clear that Big Tech’s asymmetrical influence over AI’s global production chain causes substantial distancing across its constituent nodes. The reduction of distancing across Princen’s four dimensions—geography, culture, bargaining power, and agency—constitutes a key step in curbing AI’s environmental harms, enabling greater transparency as well as a deliberate redistribution of decision-making power over the resource consumption underpinning AI’s production chain. An action framework for translators to resist AI’s ecological harms must therefore differentiate between strategies targeting industry-specific harms and strategies for reshaping collective consumption patterns related to AI.

Towards a preliminary action framework

The distancing concept is highly useful for identifying and opposing the power structures underwriting the extractivism of AI’s diffuse supply chain. However, the actions proposed in this section also take inspiration from a rich tradition of activist translation emerging primarily from the Global South (see Gould & Tahmasebian, 2020), likewise understanding activism to entail challenging power (Bandia, 2020). Moreover, these recommended actions follow Piñeiro’s (2026) emphasis on the need to connect translation and translation research with concrete acts of resistance, joining the historical struggles against extractivism that originated in the Global South and long predate AI technologies.

In light of the structural power imbalances that characterize the field, effective resistance to AI’s environmental harms will surely require coordinated, targeted actions in lieu of the atomized “footprint” ethics championed by green consumerism. As argued earlier on, this task requires not only minimizing AI’s environmental impacts within the scope of translators’ professional duties, but also actualizing the potential for translation— whether professional or non-professional—to support wider structural change. As such, resistance strategies may be categorized according to two levels or scales:

Vocational resistance denotes strategies that minimize the ecological impacts of AI language technologies deployed within the translation industry and its adjacent activities, including the use of NMT and LLMs in translation workflows as well as the customization and finetuning of these AI models for clients’ specific purposes. This perspective emphasizes translation as a profession—one that is already closely and increasingly intertwined with AI technologies.

Structural resistance denotes strategies that directly challenge the broader structural forces behind intensifying and expanding the extractive processes associated with AI’s use across all of society. This perspective emphasizes translation as a practice—one with great potential to support vital access to information and transnational solidarities.

While the resistance strategies presented here necessarily constitute a tentative and incomplete list, they may offer a starting point for translators to consider tangible actions for resisting AI’s ecological harms by decreasing distancing. These recommendations also amplify and draw inspiration from ongoing efforts by relevant activist groups.

Vocational resistance strategies

Emphasize the rejection of AI’s ecological costs as an added value of human translation. Translation buyers typically view post-edited MT output and human translation performed from scratch as competing services, with the latter understood as more expensive and time-consuming. While cost and speed are undoubtedly among most translation buyers’ top priorities, some clients may respond to acknowledgments of AI’s ecological harms by embracing AI-free translation perceived to be “ethically sourced”. For instance, organizations such as Arte es Ética4—a collective of Spanish-speaking artists, authors, translators, and many other creative professionals—advocate for the rejection of commercial AI technologies on ethical grounds, citing environmental harms and numerous other concerns.

Demand absolute transparency from AI translation/language model developers and cloud providers. Lack of environmental transparency among AI developers has become particularly acute following ChatGPT’s 2022 public release and the ensuing competitive intensification (Luccioni, Gamazaychikov, et al., 2025, p 3). Translators can collectively demand full transparency from LSPs and other commercial entities utilizing AI models regarding the precise parameters and conditions of models’ development, training, and use, as well as their estimated environmental impacts. On a wider scale, translators may organize and/or support public calls for cloud providers to provide complete information on their environmental impacts. While disclosure alone is insufficient for solving this complex issue, it undoubtedly constitutes a crucial first step in enforcing accountability.

Organize and join public campaigns to support gig workers performing underpaid data work for training AI translation/language models. One particularly vital factor to the expansion of AI is the availability of cheap data work. The Data Workers’ Inquiry5 exemplifies the radical potential of organizing remote data workers across the world—from translators to more general data annotators—in order to confront labour exploitation in the AI industry’s outsourced data work. Campaigns of this nature are well-positioned to combine calls for improved labour conditions with calls to confront AI’s environmental tolls, perhaps even discouraging translators and other gig workers from accepting jobs related to training AI translation or language models. However, this recommended action is foreseeably rather contentious: certain workers—particularly those in the Global South—face stronger socioeconomic pressures to accept these forms of gig work in the face of more drastic restrictions in their potential sources of income. Any public campaigns must acknowledge this reality.

Advocate for industry norms and contractual mechanisms that discourage or prevent the use of translators’ work as training data for training AI models. Translators’ work is often used to train AI models, meaning that the expansion of AI language technologies partially relies on this appropriation of alienated labour. The cooperative Guerilla Media Collective6 has long criticized and sought to counterbalance such extractive and exploitative practices in the translation industry. With greater bargaining power, translators could deny developers’ use of their work as training data, thereby preventing default access to training data. Translator groups could develop and advocate for standardized contracts that grant translators the power to deny LSPs or clients from using their work to train AI models. In some cases, translators may permit companies to use their data as long as they are compensated properly and any future AI models are not used for unscrupulous ends (e.g., weapons manufacturers or oil companies). This aspiration suggests a broader overarching goal of ensuring that AI models’ actual uses and benefits justify their various social and environmental costs. For now, halting the rapid and unimpeded expansion of AI infrastructures presents a more pressing objective, as this step is a necessary precursor to circumscribing AI use in light of societal boundaries. More effective and ambitious eco-translation practices may challenge the structural forces driving AI’s expanding resource consumption.

Structural resistance strategies

Use translation to promote counternarratives and create multilingual information-sharing networks for resisting data centres. Data centres’ “physicality” provides an ideal site for anchoring resistance to Big Tech’s highly diffuse and seemingly immaterial presence (Barakat et al., 2025, p. 49). Citizen actions against the construction of new data centres have already materialized in many parts of the world. Resistance to data centre construction remains highly fragmented and hyper-localized, with several notable exceptions (Lehuedé, 2022; Jansen & Havana, 2024; Barakat et al., 2025).

Jansen and Havana (2024) present the findings of a collaborative workshop aimed at discussing strategies for contesting data centres. Language, narrative, and access to information feature prominently among the report’s principles for effective resistance: the authors recommend focusing on emphasizing the stark power imbalances in data centre planning, communicating environmental harms in a straightforward and relatable manner, and framing their appeals for support around communities’ resource scarcity (Jansen and Havana, 2024, p. 12). One prominent case offers strong support for this perspective. In 2019, activists in Cerrillos, Chile, challenged Google over the company’s strictly confidential plans to construct a highly water-intensive (up to 169 litres/second) data centre there amidst the region’s severe water shortages, which already required water to be resupplied via outside transport (Ramírez, 2023, p. 61). In order to counteract the community’s information deficit, activists focused primarily on deciphering technical documentation and relaying environmental harms to community members and potential allies (Ramírez, 2023). As Hogan and Blue (2024) similarly argue, the task of organizing resistance to AI’s largely hidden environmental tolls requires linking these tolls to people’s everyday experiences. The emphasis on water—an indispensable element in everyday life, made more crucial amid regional water scarcity—can be highly effective in building local resistance to data centres and lithium mining (Lehuedé, 2024).

As such, translation efforts may broadly coalesce around delegitimizing Big Tech’s dubious claims to sustainability and community well-being in their construction of data centres. For instance, the Spanish organization Tu Nube Seca Mi Río7 is coordinating efforts with other European data centre activists to push counternarratives rooted in scientific expertise and community experiences (Barakat et al., 2025, p. 48). Campaigns may formulate narratives around “weigh[ing] the purpose of data centres and computational processes against their potential social benefits and environmental costs” (Jansen & Havana, 2024, p. 6).

In addition to pushing broader narrative shifts, activist-translator networks may take the lead in creating transnational information-sharing networks providing multilingual resources for data centre activists, including multilingual glossaries and term bases as well as community outreach materials. A strategic approach would note where data centres (planned or operational) are clustering, in order to prioritize translation into and from the languages of these areas as well as their neighbouring regions. Piñeiro’s (2026) vital work in translating materials on fracking’s socioenvironmental harms—previously available in English only—demonstrates the stark effects of language barriers in community activists’ self-education, even for languages as prominent as Spanish. Her example may serve as a blueprint for counterhegemonic translation in the realm of data centre resistance.

Provide translation in support of data centre activists’ efforts to undertake legal challenges and engage policymakers at progressively higher levels. Challenges to the legality of data centre construction and other extractive projects lacking adequate community input may benefit from multilingual research and linguistic support. For example, a coalition of Spanish activist organizations spearheaded by Ecologistas en Acción has collectively brought a lawsuit against Amazon’s cloud computing subsidiary, Amazon Web Services (AWS) (Barakat et al., 2025, p. 48). Given the technicalities of legal language, activists may find it useful to make cross-lingual comparisons of the contents of corporations’ environmental and social impact assessments (ESIAs) and similar documents supporting harmful extractive practices.

Coordinated transnational campaigns could also inspire policy changes on a larger geopolitical scale. One promising policy recommendation would be the creation of a public “database of AI uses and framework for AI’s immediate applications in order to understand the drivers of environmental impacts” (Kneese & Strubell, 2025). This database would allow the public to identify what supposed benefits actually incur data centres’ environmental costs, encouraging debate on whether these costs are justified. Another urgently needed policy is the mandatory reporting and independent oversight of data centres’ resource use and environmental impacts, preventing real harms from being “obscured through market mechanisms” by demanding “location-based accounting” (Luccioni, Gamazaychikov, et al., 2025, p. 7).

Facilitate multilingual, transnational solidarities among workers and activists at disparate nodes of the global AI production chain. It is evident that AI’s global production chain reflects significant imbalances in bargaining power between its constituent forms of labour and the firms that procure them. These inequalities lead to poor working conditions and little democratic input regarding resource consumption and subsequent ecological consequences. Translators’ multilingual competencies may be foundational to building diverse and far-reaching support among activists, workers, and others across the global AI production chain.

Essential to the AI production chain’s distancing are the pronounced information asymmetries between AI providers and those who comprise the rest of the AI production chain. Counteracting these information asymmetries will require concerted efforts to form information-sharing networks among geographically- and linguistically-diverse activists at various nodes. The AI + Planetary Justice Alliance8 is doing pioneering work in this regard, striving to form vital links between frontline communities and activists across the global AI supply chain. Translators are well-positioned to fortify such efforts, possessing not only language competencies, but also multilingual research skills and familiarity with the nuances of intercultural communication. All of these qualities are indispensable to activism and organizing among such a diverse range of workers and communities across the Global North-South axis (see Piñeiro, 2026).

Overlaps between environmental and labour advocacy amid the “contested nature of global supply chains of digitalisation” are well-established (Valdivia, 2024, p. 3). The commonalities between these spheres are not only relevant to manual labourers (e.g., cobalt miners in the DRC). In 2018, Amazon workers across the company’s operations formed Amazon Employees for Climate Justice9—recently publishing its own report condemning the tech giant’s environmental harms and sharply criticizing Amazon leadership’s fixation on expanding data centre capacity. Numerous other labour coalitions (both in the US and beyond) are working to curtail environmental harms across the AI supply chain, especially focusing on data centres (Barakat et al., 2025, p. 48).

Translation efforts may equip workers along the global AI production chain with critical knowledge regarding occupational hazards and environmental harms associated with their work, also communicating resistance strategies and opportunities for solidarity where possible. Translators may also endeavour to elevate the experiences of workers whose language repertoires generally diverge from the major languages used in international media. For instance, Barakat et al. (2025) highlight the group Securing Indigenous Rights in the Green Economy10 (SIRGE) as a “global coalition working against harmful rare earth mineral extraction” (p. 48)—an ideal candidate for external support from activist-translators.

Undoubtedly, translation of this nature is already being performed on an impromptu basis in many global contexts (Piñeiro, 2026). What a concerted network of activist-translators could contribute, in these cases, is a specialized force uniquely devoted to the vital, labour-intensive work of rendering resistance legible across languages and geopolitical contexts, uniting and compounding otherwise fragmented defiance of Big Tech’s hegemony.

Conclusion

In order for translators to meaningfully confront the ecological harms of AI technologies used both within and beyond the profession, a radical shift in perspective is warranted. As outlined in this article, a complex series of interconnected forms of resource consumption must occur before translators even have the opportunity to encounter and interact with AI technologies within the scope of their professional duties. Contrary to the AI field’s preferred narratives, the widespread embrace of individualized footprint ethics and efficiency-based technical solutions presents an inadequate response to these growing environmental tolls. As such, translators must consider structural resistance in addition to vocational resistance. Translation professionals should, of course, reflect on the ecological consequences associated with their personal use of this technology. But reducing the cumulative ecological harms of AI and its associated infrastructures will require multi-scalar efforts to resist the structural forces driving AI’s continual expansion.

This article has asserted that Big Tech’s grow-at-all-costs mandate is the primary force behind AI’s increasing environmental harms, with power consolidation among these cloud providers contributing to severe distancing along AI’s global production chain. More specifically, the scope and nature of AI’s ecological impacts remain largely concealed from public view, as the AI production chain’s various forms of resource consumption proceed according to severe imbalances in information and decision-making power. As argued here, effective resistance to these environmental harms involves reducing this distancing via strategic translation practices at both the vocational and structural levels.

The many interrelated concepts and dimensions presented here—each highly complex in their own right—have been explored in a necessarily simplified manner; this discussion merely represents a first step in a wider debate. The article has argued that contesting the rapid expansion of AI infrastructures should constitute an immediate goal in translators’ resistance to AI’s ecological harms, given the rapidly growing role of AI in the global economy and its corresponding resource consumption. But more philosophical inquiries must also question the ultimate role(s) of AI—including NMT and LLMs—in a society designed around resource consumption patterns that adhere to planetary boundaries. More practically speaking, questions related to how and whether activist-translators might receive compensation for translation work in this area have also been neglected here; the matter of (potentially) encouraging translators to produce work without pay is a serious moral issue requiring further consideration. Future work may also aspire to produce engaged research in the manner demonstrated by Piñeiro (2026), with researchers working alongside activist-translators in creating transnational solidarities across the global AI production chain.