Proposal for a Triple Bottom Line for Translation Automation and Sustainability: An Editorial Position Paper

Joss Moorkens, SALIS/ADAPT Centre, Dublin City University
Sheila Castilho, SALIS/ADAPT Centre, Dublin City University
Federico Gaspari, University of Naples “Federico II” and ADAPT Centre, Dublin City University
Antonio Toral, University of Groningen
Maja Popović, ADAPT Centre, Dublin City University

The Journal of Specialised Translation 41 (2024), 2-25

Creative Commons Attribution 4.0 International


This article is both an editorial introduction to the guest-edited special issue of JoSTrans on Translation Automation and Sustainability, and a position paper in which we propose a model for evaluating the sustainable use of automation technology in translation and beyond. As grounding notions, the article reviews definitions of automation and considers the urgency of sustainability. Thereafter we propose an adaptation of Elkington’s (1997) triple bottom line, giving equal weight to evaluation based on people, planet, and performance, describing each of these elements in turn. Finally, we introduce the articles from this special issue, in which authors describe various aspects of automation technology in translation with a focus on sustainability.


Translation technology, translation automation, artificial intelligence, sustainability, triple bottom line, ethics.

Introduction: translation, automation, and sustainability

The advent of Neural Machine Translation (NMT) with overall improved quality and (sometimes deceptive) fluency for an unprecedented number of language combinations and domains has coincided with a huge increase in translation automation. The largest free online MT provider, Google Translate, translated 143 billion words per day in 2016, and had one billion Android app user installs by 2021 (Pitman 2021). However, automation is not ‘all or nothing’, but rather varies according to the level of human input (Parasuraman et al. 2000). Improved quality and hype about its capabilities have pushed NMT, with or without post-editing (PE), into high-stakes use cases for which automation had previously been considered inappropriate (Vieira et al. 2021). Predictions that “post-editing will dominate translation production” (Lommel and DePalma 2016: 20) do not seem to have materialised in all segments of the market, with ELIS Research (2023) reporting that only 31% of surveyed European translation organisations offer PE as a product, although this percentage is growing year on year. Beyond full or light PE, there are various other modes of interaction with NMT in translation processes, such as its use as ‘just another input’ (Cadwell et al. 2016), for fuzzy match repair, and for interactive MT. There is no evidence as yet of flexibility for translators to move between these modes, as recommended in contemporary literature on human factors (such as Calhoun 2022).

In a dynamic global industry, NMT is not the only form of automation in translation workflows, particularly as more data is being gathered from translation projects. Other options for automation include error identification and correction, quality evaluation, terminology consistency checks, project management, job allocation (e.g. Herbert et al. 2023), and billing/invoicing functions. The availability of generative tools powered by artificial intelligence (AI) and using large language models (LLMs) broadens these options further, with functionalities still being uncovered in the translation of well-resourced languages (Hendy et al. 2023), with contextual awareness that surpasses NMT (Castilho et al. 2023), and improved automatic translation evaluation (Kocmi and Federmann 2023). Sánchez-Gijón and Palenzuela-Badiola (2023) found that the capabilities of ChatGPT across preparation, translation, and post-production stages of a translation project are uneven, with some workflow steps carried out successfully and others in a random and inconsistent manner. Commercial Computer-Aided Translation (CAT) tools have begun to introduce these LLM tools, often using an API1 connection to one or other provider (such as memoQ AGT, which looks to provide contextually appropriate MT based on sharing a number of fuzzy matches). The fact that ChatGPT, GPT-4, and Google Gemini were launched between the submission of articles for this special issue and their publication exemplifies the dynamic nature of technology and the relevance of the topic of automation, which is becoming increasingly crucial across translation, in educational as well as professional settings.

These developments and their repercussions within the translation industry and more broadly throughout society make it clear that narrow methods of evaluation are no longer sufficient to draw conclusions about the utility of technology or to predict the implications of its introduction. Messick (1989: 5) proposed that not only should measures be appropriate, meaningful, and useful, but that we should also consider “the social consequences of their use”, particularly if these measures or scores are used as a basis for action. The vastly exaggerated claims of human parity for Chinese-English MT in the news domain based on direct assessment evaluation had social consequences in terms of media reports that were much less nuanced than the original paper by Hassan et al. (2018). Some of us responded to suggest broader evaluation methods (see Toral et al. 2018 and Läubli et al. 2020), but a difficulty is that these do not present the reader with a simple story. Similarly, Schwartz et al. (2020) criticise a narrow focus on performance in evaluation – what they call Red AI – without consideration of efficiency and sustainability.

In this position paper which also serves as an introduction, we build on the insights provided by the papers collected in the special issue and on our individual and collective thinking while editing it, to propose a broader method of evaluating automation technology in connection to sustainability, and its utility within research and private and public organisations, using the framework of Elkington’s (1997) triple bottom line of people, planet, and profit (the latter of which we adjust to ‘performance’). It is important to state from the outset that we are not anti-technology or automation – far from it, we are all technology researchers, after all – but feel it is important to caution responsibility in its evaluation, reporting, and use, mindful of the broader consequences and high stakes. In the following sections, we present some definitions of automation and its effects on the translation industry, then explain the urgency of sustainability. Thereafter we introduce the triple bottom line and some related discussion from business ethics, and how we might practically go about a holistic evaluation of technology. To this end, taking our cue from well-established models and approaches, we propose six questions that in our view represent a useful starting point to analyse the development and adoption of translation technologies from the point of view of the sustainability of translation and automation.

The idea for this special issue originated from the long-standing common interests and collaboration of the guest editors over multiple research projects and publications in the field of human-informed translation technologies, and in particular NMT. Recent developments in generative AI lend even more meaning and importance to the key motivation for the special issue, which is to investigate the variety of novel implementations of (full or partial) automation in translation and their effects on sustainability, considering both the sustainability of the translation industry and profession as well as ecological sustainability. We discuss the contributions to the special issue prior to our conclusion (and in the video roundtable in this special issue), highlighting how they illuminate the relation between automation and sustainability of translation from multiple complementary perspectives.

Defining automation

Automation is the first grounding notion of this position paper. Parasuraman et al. (2000: 287) define automation as the “full or partial replacement of a function previously carried out by the human operator”. The corollary is that automation can “vary across a continuum of levels, from the lowest level of fully manual performance to the highest level of full automation” (ibid.: 287). Their proposed typology of automation is mirrored by many others in the literature, according to a meta-analysis by Vagia et al. (2016). The definition of automation as replacement pits humans against machines, which O’Brien (2023) criticises as an antagonistic dualism that does not afford the potential automation of tasks that were not previously performed by humans.

In translation, few if any translators now receive absolutely no machine assistance, with basic spelling and grammar correction available in most text editing interfaces at a minimum. As we move through the automation typology, users are offered several options (as may happen within a CAT tool) or just one (as may happen when post-editing). As the machine begins to assume more control, humans may have the opportunity to veto or approve the choice (as happens with interactive MT) or may be not involved at all (as in raw MT). However, it should be noted that humans are still involved in the development, curation of data, and implementations used for raw MT processes, even if in a somewhat less visible, subtle manner (Gaspari 2022).

This sort of scale was applied to translation by Hutchins and Somers (1992) during the era of rule-based MT, who influentially included human-aided MT and machine-aided human translation as transitional areas in a translation automation continuum. Christensen et al. (2022: 35) update this model to create a six-point typology of translation automation along the lines of the Society of Automotive Engineers’ classification of driving automation, noting that while these roughly align, they also differ in that society has zero tolerance for driving automation errors that will have fatal repercussions, whereas “it seems that society has already grown accustomed to imperfect translations”.

In these typologies, the automation level is usually static. Lommel’s (2021) proposal for responsive MT suggests a dynamic interaction, with automatic domain adaptation, contextually aware MT systems that are adaptive based on feedback and that automatically match with application and usability requirements. A dichotomy that appears in the automation literature is between adaptable (human-controlled) and adaptive (machine-controlled) dynamic switching between automation levels. Vagia et al. (2016) suggest that the former can avoid “some of the pitfalls” of the latter, echoed by Calhoun (2022) who finds improved situation awareness and perceived control on the part of users who can set their desired level of automation.

Way (2013: 2) examines the key variables that influence the feasible or desirable levels of automation in (machine) translation along with some case studies, arguing that the “degree of human involvement required – or warranted – in a particular translation scenario will depend on the purpose, value and shelf-life of the content.” More recent work (e.g. Rothwell et al. 2023) added risk to this guidance, following Canfora and Ottmann’s (2018) typology of translation risk, ranging from the risk of miscommunication to the risk of injury or death. This would seem to still hold true for generative AI tools. Of course, risks to sustainability as detailed in the next section are at a higher level than those originating in the texts themselves.

The urgency of sustainability

According to the Brundtland Commission Report (1987: 16) sustainable development “meets the needs of the present without compromising the ability of future generations to meet their own needs.” While these early notions of sustainability mostly related to environmental sustainability, this has broadened over time to include stewardship of our social environment, as evidenced by the United Nations Sustainable Development Goals (SDGs; UN General Assembly 2015)2.

Two major news stories at the time of writing regard an unprecedented heatwave in many parts of the globe and a prolonged strike by screen writers and actors in the United States. The former follows alarming research (Thompson et al. 2023: 1), predicting that many vulnerable populations will likely be subject to more frequent extreme heat due to climate change, and will be “increasingly exposed because of limited healthcare and energy resources”. Stoddard et al. (2021: 654) cite “narrow techno-economic mindsets and ideologies of control” as some of the reasons why three decades’ worth of climate mitigation efforts have been ineffective. The Hollywood screen writers’ strike may appear unrelated, but this was also triggered by narrow techno-economic mindsets and ideologies of control, as writers are forced to argue for project-length contracts and try to prevent their output being used to train AI systems that might eventually displace them.

Both ecological and workplace sustainability are core issues in translation and have been for some time. As Cronin (2019: 516) writes, translation is “inevitably implicated in any discussion of what happens to technology in an age of accelerated climate change”. Our engagement with technology needs to be situated “within the carrying capacity of a planet with finite resources and an ever shortening timeline of climatic viability” (ibid.: 516). Yamada’s (2023) more modest aim is to use translation process research to demonstrate that human translation and PE are not so dissimilar, arguing that the devaluation of PE threatens the sustainability of commercial translation. As we shall see in Section 4.2, there is ongoing work in AI and sustainability, with a growing realisation of environmental harms beyond carbon emissions, but efforts to improve efficiency are currently outpaced by the growing size of LLMs as the most effective way to increase output quality.

UN SDG number 8 proposes “sustained, inclusive and sustainable economic growth, full and productive employment and decent work for all”, but a focus on reducing labour costs by imposing automation has affected many industries, including translation. Moorkens (2020) notes that a tendency to atomise some parts of translation work is incompatible with findings on motivation and satisfaction, in that positively motivating factors are often intrinsic, involving meaningful work with recognition and responsibility (Herzberg 1976). According to Docherty, Kira, and Shani (2008: 4), instead of focusing on “short-term, static efficiencies such as productivity and profitability; we must also focus on long-term, dynamic efficiencies such as learning and innovation”. This goes against long-standing business orthodoxies, such as Porter’s (1980) five forces, which advise against allowing suppliers or customers to form an allegiance in order to limit profits by increasing costs or limiting prices. It is with work such as Porter’s in mind that the move to digital translation platforms with controlled methods of communication between translators who work on decomposed portions of text, as parts that may be considered interchangeable, makes business sense, at least in the short term (Agorni and De Bonis 2022).

Triple bottom line

Elkington (1997) was not the first to propose that environmental and social benefits should be of equal importance as economic ones, but his idea of the ‘triple bottom line’ of people, planet, and profit has stuck. The ‘people’ in this model are not only workers, but also other stakeholders in an organisation. Following Stakeholder Theory (Freeman 1984), translation stakeholders might include translators, target text end users, institutions and citizens, trainers and educators, members and representatives of professional associations/bodies, shareholders, company owners, project managers, and workers in all areas of translation. We could also include people living in the area where business is conducted or journalists who report on an industry, who Phillips (2003) defines as derivative stakeholders, who have an indirect connection in the organisation. According to the triple bottom line, an organisation should not exploit these stakeholders, and should ideally contribute to their wellbeing, along with all others involved in the production of their output.

The ’planet’ part of the model involves not only causing no ecological harm and limiting energy consumption and waste, but also carrying out a full life cycle assessment of products. Brevini (2022) highlights not only the power and water requirements to operate the data centres required for cloud computing behind contemporary AI, but also the rare earth minerals required to build machines and the e-pollution caused by the disposal of outdated hardware in huge dumps, such as those in Kenya and Cambodia. As Cronin (2020: 520) writes, “there is nothing immaterial about the material consequences of virtual technologies”. Williams (2011: 355) details the risks to people in the manufacture, disposal, and recycling of technology, such as “exposure to ancillary chemicals used in high-tech processing, in particular making semiconductors”, including known carcinogens.

The argument put forward to support the triple bottom line is that environmental sustainability can be more economically profitable for an organisation in the long run (Elkington 1997: 38). Melé (2009) believes that running an organisation in accordance with the common good – for the good of our larger community – should be a successful business approach and does not mean ignoring the needs of shareholders. For Melé (2019: 298), sustainability is “nothing other than the common good for future generations”. This fits with Elkington’s (1997) view of sustainability as a new form of value that society will demand.

The ’profit’ part of the model may seem obvious, but Elkington (1997: 74) expands this to focus on the economic sustainability of an organisation, including not only economic capital, but also human capital (“a measure of the experience, skills, and other knowledge-based assets of the individuals who make up an organization”), intellectual capital, social capital, and natural capital. This view of profit incorporates the real economic benefit enjoyed by the host society and, more broadly, ecosystem. The corresponding concept to the traditional bottom line in business of ‘profit’ when considering technology is ‘performance’. Schwartz et al. (2020: 62) write that the “push to improve state-of-the-art performance has focused the research community’s attention on reporting the single best result after running many experiments for model development and hyperparameter tuning” at the expense of improving model efficiency. The tendency is to evaluate the latest (AI-based) technologies based on a narrow view of performance, as we shall see in Section 4.3, ignoring inconvenient elements, such as biases in output, that are paid considerably less attention. Of course, we recognise that we cannot ignore the bottom line of performance or quality, but we suggest that we can, and should, report it alongside the implications for people and the planet.

The triple bottom line model was intended to encourage long-term thinking, a broader view of the purpose of a business or organisation within society, and a move away from simplistic short-term answers, from a single bottom line that looks only at immediate economic profit. As Elkington later wrote (2018: 8) the “stated goal from the outset was system change — pushing toward the transformation of capitalism”. Accordingly, our suggestion to use this model for a broader and more truthful evaluation of translation technologies and their consequences moves away from simplistic accounts that tell a deceptively straightforward story towards a more complex and nuanced representation of automation that moves beyond efficiency and short-term gain.

4.1 People

The first consideration for people as users of translation technology is quality, and when quality is decoupled from users, for example by using automatic evaluation metrics that do not correlate well with human judgement, the use of those metrics should be called into question. However, in this article we link quality to performance, as explained in Section 4.3. Reduced translation quality will introduce risk to users, but a key element of the triple bottom line is that each part is crucial and interlinked. As Abdallah (2014) proposes in her three-dimensional model of translation quality, the social quality of translation, asking who does what under what circumstances, is intertwined with the quality of the process and of the final product.

People are also the source of training data for contemporary machine learning technologies (Gaspari 2022). Translators may have contractually agreed (or not) to allow their work to be repurposed for MT system training, but the use of webcrawling for data acquisition is currently standard, without any real legal basis. Some artists have begun to use cloaking tools to prevent their work being used as training data (Shan et al. 2023) and website administrators may use the robots.txt file to try to prevent crawling of their online data (if the crawler has been set to respect this instruction). As human data is the gold standard, and machine-generated data begins to appear frequently on the internet (causing deteriorating quality if used for training machine learning systems) human data and “data about human interactions with LLMs will be increasingly valuable” (Shumailov et al. 2023, 2). Organisations will want to avoid the “ouroboros effect” (Moorkens 2023a: 18) of systems being trained on their own output, while being cognisant of the rights of people not to have their data used without consent.

Translators and workers within digital platforms often have little choice in whether or not their data is collected and used for MT system training. Their work may even be intended solely to produce data for training or to provide human feedback in order to reduce bias or harms in system output (Ouyang et al. 2022). Workers within digital platforms are also often prone to algorithmic management, with little or no human oversight, and no explanation or accountability for decision-making. While this can reduce unpaid work in searching for jobs within a platform, there is the necessity to conform to what the system rewards, with timeliness and polite communication replacing the need to maximise translation quality, and translation norms being displaced by algorithmic norms (Moorkens 2023b).

The increased automation within the translation industry, especially within digital platforms (used regularly by 89% of translator respondents to a survey by Pielmeier and O’Mara (2020)), presents a risk to organisational sustainability as mentioned in Section 3. According to Wheelen et al. (2018), companies that focus on business sustainability are rewarded by reduced staff turnover and increased employee effort. They also find that companies that focus on cost, replacing workers with automation where possible, demonstrate the opposite effect. The recommendation is thus to use automation to add value or to diversify the business offering, automating what was not being done before, and making sure to foreground satisfying, motivating work. A sustainable work system is “aimed at the regeneration of the resources it utilizes – human, social, material, and natural resources” (Docherty et al. 2008: 4), and the balancing of these and the needs of various stakeholders is difficult and will require constant evaluation and recalibration. These natural resources are where we turn to in the following section.

4.2 Planet

The most obvious concern for environmental sustainability is the energy requirement for automation technologies, sometimes referred to as compute costs. The contribution by Strubell et al. (2019) raised awareness of the carbon emissions related to training machine learning models based on the Transformer neural architecture (by far the most common architecture for MT and text generation in general at the time of writing; Vaswani et al. 2017), with the graphical processing unit (GPU) emissions for training a large model equated to the output of 1.5 cars over the 20-year lifetime of those cars, without even including the power and cooling requirements for the whole computer, hence their recommendation to “prioritize computationally efficient hardware and algorithms” (ibid.: 3646). Luccioni et al. (2023) also highlight the additional emissions related to generative AI when compared with task-specific systems, particularly at the training stage.

Wu et al. (2021) include the manufacturing and operational costs of equipment in their calculations and report 25% increases in efficiency of machine learning models over a two-year period. There has also been a trend in recent years towards pre-trained models, rather than always training from scratch (see Doğru and Moorkens in this issue), and researchers such as Jooste et al. (2022) have examined ways to make machine learning systems more efficient. This promising work should, however, be placed in the context of increasingly large language models. Shterionov and Vanmassenhove (2022) highlight the difference in emissions based on energy sources and location, with training in Ireland utilising more renewables than fossil fuel-based power in the Netherlands. Dodge et al. (2022) also add consideration of the time of day for training and the data centre location.

The manufacturing and disposal risks to workers in particular, and more broadly to citizens, are part of a larger sustainability problem across the life cycle of technology as detailed early on by Williams (2011), well before the latest AI-led technology breakthroughs. Problems include the release of toxins in incineration, release of harmful metals and compounds when computer equipment is (inadvertently) mixed with general waste, and burning of some materials within informal recycling processes. Due to high labour costs and environmental restrictions in many countries, these informal processes tend to be pushed to developing regions for reasons of cost effectiveness and, often, more permissive laws. Again, we have no easy answers as to whether a technology is worth developing or not, or – more specifically – under what circumstances it could or should be developed, balancing its benefits and costs. We would, in principle, agree with Williams (2011: 357) who concludes that “[u]nderstanding the interaction of ICTs with economic and social systems presents significant and interdisciplinary methodological challenges”.

Ethical best practice in research has moved on from the edict to ‘do no harm’ to actively do good. Antonopoulos et al. (2020) explain how machine learning technology can automate demand response in energy providers, routing energy quickly and effectively, managing costs and resources. This forms part of the wide-ranging proposals for machine learning to improve sustainability by Rolnick et al. (2023): ideas include modelling emissions and forecasting demand in electricity provision, reducing the need for standby generators, using sensors to reduce waste and harmful emissions, improving transport efficiency, optimising farms and industries, and (perhaps most ambitiously) modelling and managing emissions at a global level. While some of these ideas are currently fanciful, we would argue that those with concrete impact should be examined and balanced against the immediate negative impacts.

4.3 Performance

The traditional bottom line for technology is performance, which in translation is quality. As detailed in Castilho et al. (2018), there are many ways to monitor and evaluate translation quality, and these may be appropriate for some products, processes, and scenarios but not for others. Quality is “never absolute but depends on both context and situation” (Drugan et al. 2018: 42). There is undoubtedly a place for automatic evaluation where human evaluation is too slow or expensive. What is key is not to overgeneralise based on a narrow set of results, for example using sensationalist terms such as the already mentioned ‘human parity’ or ‘superhuman translation’ based on a small-scale evaluation of translation in a single language pair and domain where not all independent variables are given due consideration and controlled properly (Läubli et al. 2020). Researchers such as Freitag et al. (2020) and Kocmi et al. (2021) advise against the use of BLEU (Papineni et al. 2002), the most popular automatic evaluation metric in the MT field during the last two decades, on the basis that its overuse has impeded MT development. However, the best available metrics show weaknesses too, such as for named entities or numbers (Amrhein and Sennrich 2022). This is an area of active research, with new automatic evaluation metrics being regularly proposed (e.g. Fernandes et al. 2023; Guerreiro et al. 2023) and new standards for translation evaluation due to be published by ASTM and ISO3.

Reduced translation quality could not only introduce risk, but could also push the effort in comprehension towards the end user, as already noted by Pym (2012). There may be problems with mistranslation, along with problems of hegemonic, gender, or racial bias for the reader to disentangle. However, we must also be mindful that quality needs differ based on circumstances, as argued by Way (2013). For example, in crisis settings, the “gold-standard expectations enshrined in ISO standards and codes of conduct seem unattainable” (Federici and O’Brien 2019: 11), but one can argue that in critical situations an imperfect translation is better than no translation at all, especially if it meets some urgent and serious needs of people in danger without causing harm. We might usually want excellent quality translation in general, but at times reduced quality translation is better and less risky than none at all, and may in fact be the only viable or possible option for very solid reasons.

If a (translation) technology is to be useful, there will have to be a clear benefit in performance or an economic motivation to its deployment. However, this is where gains in efficiency need to be placed in the context of the human, social, and environmental spheres. These are not to be played off against one another, but to be considered in the round and in how they contribute to overall sustainability. Elkington (1997: 316) stresses the need to change from an extractive approach to “modes which, over time, actively rebuild economic, environmental, human and social capital”. In business, the COVID-19 pandemic demonstrated the lack of resilience in ‘just in time’ supply chains that foregrounded efficiency above all else (Remko 2020). Similarly, a short-term focus on performance and cost without a similar corresponding focus on sustainability is unlikely to lead to real, long-term benefits. By all means, efficiency in comparison with benchmarks is important, but not at all costs.

Six questions for modelling a triple bottom line for translation automation and sustainability

Putting this technology triple bottom line into practice is more complex and involved than carrying out a single static measurement of the quality of a tool or technology. Rather, it should involve continuous evaluation and recalibration of the technology and its deployment. In this section, we propose six questions – two each on people, planet, and performance – that could form a useful basis for analysis before and after the development and introduction of a technology, which we feel are particularly relevant to translation technologies in the context of automation and sustainability in the age of AI, as per the focus of this special issue.

Who are the stakeholders and how does this technology affect them?

The most obvious stakeholders in translation are source text authors, translators, clients and end users. How will this technology change translators’ work and if it does, will this be positive or negative? What will be the impact on clients? Will the text or product for the end users change for the better or worse, and might it expose them to risks? What about other stakeholders “to whom the organisation has a moral obligation, an obligation of stakeholder fairness” (Philips 2003: 30) such as the people who create or provide the data, the company that develops or uses the tool and its employees? Will they be treated more or less fairly as a result of this technology? How will society be affected more broadly? There is an argument that the use of AI technologies might have broader repercussions on society, or that overuse of MT might negatively affect a minoritized language (see Aranberri and Iñurrieta in this issue).

What are the consultation, training, and feedback needs when introducing this technology?

Cadwell et al. (2018: 317) argue that “translators ought to be included in the change process from the very beginning” when introducing MT. This would seem to be appropriate for technology more broadly, as introduction via discussion and ideally based on consensus is more likely to be accepted than unilateral imposition of a tool, which might lead to tension, if not outright rejection. Cadwell et al. (2018: 317) highlight the importance of agency here, not giving the impression that the technology has taken precedence and is “inevitable, no matter how unfitting it might be for the task at hand”. Bywood et al. (2017), Vieira and Alonso (2018) and others have also stressed the importance of training, but there are many ways to carry out training (see Bell et al. 2017) and it needs to be carefully planned. Cadwell et al. (2018: 317) also recommend that translators “monitor and improve their own quality and productivity in an engaged way, without feeling externally monitored”. Ongoing evaluation and recalibration should lead to improved and more effective interaction between people and technology.

What is the environmental impact of developing and using the technology?

How can it be mitigated or ideally offset?

Measuring energy requirements is difficult, with GPU time possibly the most straightforward way to approximate the energy consumption of a machine learning technology. As Shterionov and Vanmassenhove (2022) note, emissions will vary depending on the energy source, e.g. a cloud tool hosted in Iceland using renewable energy will emit less carbon dioxide than one hosted in Malta or Luxembourg4. Will the technology save energy elsewhere? During the COVID-19 pandemic, many meetings became video conferences, facilitated by cloud-based tools hosted in data centres. This also meant that participants made fewer journeys. However, the overall costs and benefits are very difficult to measure and compare against each other, also considering the shifts in gains (and losses) across individuals, companies, organisations and institutions.

Is the quality provided by the automation technologies appropriate for the users and purpose?

Is performance efficient and consistent?

Koby and Melby (2013: 178) write that a quality translation “demonstrates required accuracy and fluency for the audience and purpose and complies with all other negotiated specifications, taking into account end-user needs”. There are lots of ways to measure translation quality, both automatically and using human evaluation (see the contributions to Moorkens et al. (2018) for many examples). Additionally, if a technology is slow or inconsistent, it will produce erratic output and frustrate users. What is important is that the evaluation method is appropriate for the users and purpose. We also note that in the MT field, evaluation does not normally take the translation brief into account.

Providing answers to these questions that we propose here will give a far more nuanced picture of a technology than a static measure of performance such as a BLEU score. However, we are aware that benchmarking, comparing, and communicating this holistic result will be difficult, introducing new challenges in a world focused on short-term easily defined gains.

The contributions to this special issue

Following this position paper, nine articles in this special issue address different aspects of translation automation and sustainability, beginning with four contributions that focus on PE. Rico Pérez revisits PE guidelines, arguing that the conventional distinction between light and full PE is no longer fit for purpose due to the generally improved quality of state-of-the-art NMT. She proposes a redefined and replicable set of PE guidelines to mitigate factors that may induce tension and to discourage negative attitudes towards PE.

Dai and Liu investigate the impact of source-text readability on PE effort for English-Chinese NMT. Their results show that readability has a significant effect on cognitive effort, and that readability metrics can predict PE effort to a certain degree, although no single formula was able to predict all of the effort indicators, suggesting that a combination of metrics may be useful for effort prediction. Nitzke et al. revisit their previously published decision tree (Nitzke, Hansen-Schirra, and Canfora 2019) to aid decisions about a project’s suitability for MT or PE. Based on interviews with 19 stakeholders, the authors reduce their model to four main decisions concerning the suitability of the source text, the reliability of the MT output, the purpose of the final text, the quality requirements for the final text, and social sustainability within translation workflows.

Guerberof-Arenas, Valdez and Dorst’s study engages master’s students of translation at two Dutch universities, who translated and post-edited literary texts in English at the beginning and at the end of taking translation technology modules. The study shows that the students tended to be more creative when translating unaided, although they made significantly fewer errors in PE, especially at the start of the training, than when translating manually. The study provides valuable insights on the complex connections between PE training, creativity, and translation proficiency.

The following two articles relate to subtitling. Guerberof-Arenas, Moorkens and Orrego-Carmona’s more than 200 English-speaking participants were surveyed on their narrative engagement, enjoyment, and translation reception of Latin American Spanish to English subtitles, comparing MT, PE, and human translation. The largest disparity in translation reception is between MT and PE, but the study also reveals that achieving publishable PE subtitles requires a substantial number of edits, as indicated by high HTER scores. The authors emphasise the importance of considering factors such as time and remuneration for PE tasks, to ensure acceptable quality and sustainable work processes.

Tamayo and Ros Abaurrea analyse speech recognition software for the intralingual subtitling of news programmes in Basque. Their evaluation identifies room for improvement in recognition, particularly regarding the handling of punctuation, the recognition of proper nouns, and speaker identification. Despite these weaknesses, Basque speech recognition seems promising, considering that speech recognition is still in the early stages of development for a low-resourced language like Basque. Aranberri and Iñurrieta also consider language sustainability for the Basque-speaking community, surveying translators and interpreters, language professionals, and general users to investigate their current and expected use of Basque-Spanish MT. Respondents are generally positive about the future for MT and Basque, with some concerns expressed about the impact that MT might have on the quality of published Basque texts, particularly administrative texts originating in Spanish.

The final two articles relate to translation workflows. Doğru and Moorkens investigate the impact of data augmentation using TMs for desktop MT fine-tuning in OPUS-CAT, assessing the utility of desktop MT for professional translators by fine-tuning MT engines in three language pairs (English → Turkish, English → Spanish, and English → Catalan) with localization corpora of varying sizes. The results demonstrate promising improvements in translation quality across all three language pairs, underscoring the potential of desktop MT applications to deliver high-quality translations while offering benefits such as privacy, confidentiality, and reduced computation power usage, the latter thanks to the use of pre-trained MT models. Finally, Silva et al. revisit the Multidimensional Quality Metric (MQM) evaluation framework, originally oriented towards European languages, and propose an amended error typology to better suit East Asian languages such as Mandarin, Japanese and Korean. They also propose a Quality Estimation (QE) method to predict the MQM scores of MT outputs at scale, showing a fair correlation with the human judgement.

Conclusions and avenues for future work

The work of Elkington (1997: 141) and others has been influential in changing mindsets in business, but as he wrote in 1997, “the eco-resource challenge may prove to be the relatively easy part of the sustainability transformation, while the socio-economic challenge looks likely to be more intractable”. A broader notion of sustainability will be important so that people are not left behind. This is the thinking behind the UN SDGs, even if these might compare unevenly (as noted by Buts et al. 2023). Sætra (2021: 16) argues that we need to look at AI in context and that “doing so shows that it is intimately tied to severe threats to most of the SDGs”. If a triple-bottom line evaluation of translation technology is to be effective, it needs to be able to identify routes of research and development that are not sustainable and thus should not be pursued.

In revisiting the triple-bottom line, Elkington (2018: 10) sees radical intent as necessary to spur “the regeneration of our economies, societies, and biosphere”. This is inarguably more important now than ever before and can only become more urgent, particularly with the young generations appearing to take a keen interest in these issues, especially with regard to environmental concerns. While sustainability as a topic has rarely been directly and specifically addressed in translation studies research, this special issue, along with pioneering work by Cronin (2017), the special section edited by Buts et al. (2023) and other recent work addressing sustainability (e.g. Todorova 2022) shows that researchers in translation studies are interested in and concerned about sustainability. However, as discussed in Section 1, sustainability does not appear to have been the foremost consideration in the development and deployment of automation technologies in translation. The model proposed in this position paper may not be the single best way forward for evaluation of automation technologies, but it is increasingly clear that it is time for new ideas and that the old methods that focus only on isolated measures of performance need to evolve and be enriched.



A person in a black shirt Description automatically generated with medium confidenceJoss Moorkens is an Associate Professor at the School of Applied Language and Intercultural Studies in Dublin City University (DCU), Challenge Lead at the ADAPT Centre, and member of DCU's Institute of Ethics, and Centre for Translation and Textual Studies. He is General Co-Editor of Translation Spaces with Dorothy Kenny and coauthor of the textbooks Translation Tools and Technologies (Routledge 2023) Translation Automation (Routledge 2024).

ORCID: 0000-0003-0766-0071


A person smiling at the camera Description automatically generatedSheila Castilho is an Assistant Professor in SALIS at Dublin City University. She worked as an Irish Research Council Research Fellow at the Adapt Centre on the DELA Project, which involved testing sentence-level metrics for document-level machine translation evaluation and establishing best practices. Sheila has actively contributed to various EU projects. Her research output includes over 40 publications, covering topics on translation technology, post-editing of MT, user evaluation of MT, and translators' perception of MT.

ORCID: 0000-0002-8416-6555


Ein Bild, das Menschliches Gesicht, Person, Krawatte, Kleidung enthält. Automatisch generierte BeschreibungFederico Gaspari teaches English language and translation at the Department of Political Science of the University of Naples “Federico II” (Italy) and collaborates with the ADAPT Centre of Dublin City University (DCU, Ireland) on EU-funded international research projects. His main research interests include language and translation technologies, applied English linguistics, corpus linguistics and corpus-based translation studies, and he has published widely in these areas.

ORCID: 0000-0003-3808-8418


A person with glasses and beard wearing a vest and vest Description automatically generatedAntonio Toral is an Associate Professor in Language Technology at the University of Groningen, where he coordinates the Computational Linguistics research group. His research interests include the application of machine translation (MT) to literary texts, MT for under-resourced languages and the computational analysis of translations produced by machines and humans. Prior to starting a faculty position, he was a postdoctoral researcher and research fellow at Dublin City University, and before that a PhD student at the Universitat d'Alacant and at the Istituto di Linguistica Computazionale. He coordinated the Abu-MaTran project, which was flagged as a success story by the European Commission and won the best paper award at MT Summit 2019 for his work on post-editese.

ORCID: 0000-0003-2357-2960


A person smiling at camera Description automatically generatedMaja Popović is a Research Fellow at ADAPT Centre at the School of Computing in Dublin City University (DCU). Her main interests are machine translation, multilingual natural language processing, as well as human and automatic evaluation methods for NLP. She has over 80 scientific publications in book chapters, journals, conferences and workshops. She is associate editor of the LREV (Language Resources and Evaluation) journal.

ORCID: 0000-0001-8234-8745



  1. Application Programming Interface.↩︎
  2. See Sætra (2021) for an analysis of AI using Sustainability Development Goals.↩︎
  3. These are expected to appear at and after publication.↩︎
  4. See↩︎