The language industry, automation, and the price of finding the right words

When Korean speakers denounced that Netflix’s ’botched’ English-language closed-captions for Squid Game change the show’s meaning about capitalism and class struggle, viewers wondered why streaming giants are not trying harder to find the right words. But the survival drama has also come under scrutiny for the conditions under which much language work in the creative sector is performed: fast turnaround, poor professional acknowledgement, and pay as low as 75p per minute of programme time. These conditions reflect widespread industry norms. Add that streaming giants are implementing artificial intelligence (AI) that can pre-process and translate subtitles into multiple languages, stir thoroughly, and you will get why language professions have a hard time maintaining their boundaries.

The language industry is among the fastest growing industries, with addressable global market of USD 23.8bn in 2020 and hiring activity which has climbed 65 points since 2018. The UK is the largest single-country market for language services (GBP 1.5bn). Figures from the European Language Industry Survey 2021 show strong industry structure; 49% of its segments are worth up to 100M €, despite slower growth during the COVID-19 pandemic. Recently, the language industry has expanded considerably towards the gig marketplace, up to 20 percent value in 2020-2021.

More interestingly, the industry is a gripping reflection of the evolution of work in the knowledge economy, based on the production and consumption of intellectual capital. As microcosmic of larger trends of work, it is therefore a useful point of discussion. Among the many sectors revealing the interrelation between technology and labour, the language industry is in fact one of the most susceptible to automation. What was previously high-skilled work dependent mainly on self-employment and on the supportive use of technology, encompassing training and qualifications, ISO standards and accreditation schemes, is now rapidly growing as a form of precarious work subject to extreme subcontracting and quick digital change. It is important to focus on how some dynamics generally linked to contemporary work and technology are affecting the language industry, fuelling the misperception of language services as low-value labour and commodities.


From outsourcing to gigging?

Microeconomic analyses argue that since the 1970s the language industry has been an example of the post-Fordist regime through elements as outsourcing, automation and workers’ flexible specialisation, which have also informed the industry’s race towards the platform economy.

For most other sectors, language services such as interpreting and translation do not fall among organisations’ key activities, core strategies, or even competencies. As the Forbes Business Council has noted, while these services allow companies to expand their reach and position themselves as global enterprises, they are not needed daily. They are also costly activities which do not directly generate revenue. For organisations it is easier and more efficient to subcontract and outsource the required services to external providers. This ingrained externalisation is mirrored by the industry’s commonplace contractual bases. Only 6% of language professionals enjoy permanent jobs as in-house or ‘staff’ employees, e.g. the interpreters hired by the EU institutions or UN agencies. Dependent self-employment, whereby professionals are contracted by language service providers (LSPs) to work for a third party alongside working for direct clients, is pervasive and accompanied by job stress factors such as the ‘feast or famine’ nature of the work, unpredictable income patterns and lack of labour rights.

On the other hand, it is a truism that technologies such as the Internet, computer-assisted translation tools (CAT) and translation memories (TMs) have always characterised the industrial landscape of language work. For instance, translators and interpreters have relied upon computers for tasks as terminological research for many years. The more recent adoption of AI and digital platforms, nonetheless, has scaled up processes such as crowd-working, concurrent translation and agile localisation processes. Both the outsourcing-friendly orientation of the language industry and a tendency towards agile work have pushed language work to develop also in the platform economy. Such model is becoming normalised as a large scale-production ecosystem, to the point that scholars have defined it as the ‘Digital Taylorism’ and the ‘uberisation’ of the language industry. Currently, there were more than 55 marketplaces and platforms tailored specifically to the mass-production of language services aided by technological innovation. How is this integration faring?


Automation and polarisation

A commonplace aspect of the capitalist workplace is that technology is deployed to both reduce costs and control labour which, in turn, result in the de-skilling of workers. Another common argument revolves around the polarisation of skills, where labour splits into jobs at the bottom requiring lower skill level, and those at the top, requiring greater skill level. The third argument usually refers to upskilling, the process of enhancing skills in a certain area. As previous contributors to Futures of Work have discussed, the future landscape of work is shaped not only by ideas of technology, but rather of automation, with machines matching or outperforming humans in a fast-growing range of tasks. The activities most susceptible to automation are routine cognitive tasks like data collection and data processing, as well as routine manual activities in predictable environments. What is left for human workers is labour tied to intellectual and creative skills that are hard, if not impossible, to reproduce through an algorithm.

Let us take illustrative examples from two language industry segments. In the translation and localisation segment, increasingly AI-powered platforms such as Unbabel use machine translation and automatic quality estimation to translate texts (‘units’), only distributing post-editing tasks to workers (‘agents’) for word alignment, syntax parsing and spellcheck. This translation pipeline “streamlines and scales globally without compromising quality” and only in some cases pairs its “invisible translation layer” with “human-in-the-loop quality enhancement”. Axiomatically, such automation practices diminish rather than increase language professionals’ skill use. As the Unbabel  example shows, AI carries out the core of the translation work and leaves the individual to fill its gaps. Thus, professionals’ autonomy and decision-making over linguistic-cultural choices is lowered, regardless of the fact that these are key to their know-how.

The interpreting segment so far has better resisted automation. Interpreting requires constant cognitive exercises in decision-making as spoken or signed language is uttered in an ever-moving flow. Interpreters convey messages as they are being shaped while accommodating for speech components, half-baked or complex sentences, emotions, and cultural references. Since interpreting does not align with just ‘replacing’ words and happens in real time, its automation seems far-fetched. Notwithstanding, research exploring the R&D of language technologies suggests that ‘machine’ interpreting is to be taken seriously. Due to the widespread use of remote interpreting (services performed via online conference tools), digital data as oral speech segments and their multilingual equivalents are stockpiling. This data is feeding deep learning technologies and natural language processing (as Google Translate or iTranslate Voice) that can code and automate speech.

The future of the industry appears moving to a U-shaped polarisation in employment as a factor of skill level. Most of language professionals’ activities seem likely to shift toward a de-skilling move, such as post-editing output. On the other hand, they are still commissioned to work on content and tasks defined as ‘creative’ that play a significant function in communication, while also being expected to tune to the automated systems. However, the state of the industry implies (as in the example of Unbabel, which is the norm rather than the exception) that intellectual work is not necessarily factored in the translation pipeline. Techno-optimists emphasise that adaptive experts will still hold onto their intellectual work and differentiate themselves from automation and less qualified individuals, hence dictating labour price and autonomy over their work. Regardless, while the demand for language work worldwide is growing as one of the main inputs of the knowledge economy, much of such work is already stuck at the lower end of the market, where automation and cheap labour are used as major selling points. As a recent McKinsey Global Institute study shows, there is a negative correlation between tasks’ wages and required skill level on the one hand, and the potential for their automation on the other. Automation reduces demand for low- and middle-skill labour in predominantly routine, low-paying tasks, while increasing demand for high-skill, high-earning labour that requires problem-solving skills. Simply put, automation is skill-biased. As the example of the language industry implies, this polarisation should urge a large-scale move towards empowering a more equitable and sustainable language services ecosystem.


Diversification and fragmentation

The logical consequence of automation is captured by treating language labour as an economic input to be sliced and dispensed as required. On Netflix, content can be made available to more than 190 countries in less than 34 hours by dividing video source files into chunks processed by more than 200 translators at once. This dynamic is accompanied by a vast fragmentation in job profiles and activities and is particularly strong in the platform economy. Recent research by language industry intelligence organisation Slator reports, for instance, that there are 600 overlapping job titles related to skill sets covering macro-areas such as ‘translation’. Multimodal translator, transcreator, pre- and post-editor, terminological expert, or communication consultant are just some of the overlapping profiles that permeate the market. Those titles are grouped by function into top horizontals — e.g. translation, localisation, audiovisual content creation, marketing, language engineering — and into verticals by end-customer industry — such as life sciences, finance, media and government.

As online-based work ramifies, language workers diversify their job titles accordingly across digital marketplaces to emerge in clients’ online searches and be noticed by potential hiring firms. This jobscape suggests diversification as a function of securing tasks, since it also goes hand in hand with work fragmentation. Language specialists, e.g. translators, used to manage a project from start to finish. Now projects are often decomposed into small chunks and posted online, as on LSPs’ interfaces or on platforms, to maximise cost and efficiency in return for micropayment. Securing tasks (and income) can be more easily achieved by fitting into one’s profile as many compatible areas and skills as possible, such as translation and post-editing. This strategy also helps meeting the demands of algorithmic control. Algorithmic control is a set of specifications acting as ‘virtual automated managers’ which coordinate and monitor workers’ activities. In most cases, algorithmic control makes some profiles more or less visible in client search, allocating work in a process of real-time digital surveillance based on skill set, productivity and price offered. These cloud-based iterations of language work require thousands of workers in an open, global marketplace to deal with a heavily siloed, automated supply of work, only to get paid less than the average wage.

Such ramifications are dangerous as, in the near future, they might fuel a full ‘gig language economy’ model, where much workers’ skill diversity and quality production go undetected and under-paid. These practices may reduce language professionals to exchangeable tools, swapped out without any detrimental effect to the system as a whole as a result of algorithmic control and the flattening in the demand and supply of skills within a ‘pay as you go’ and vastly automated system. The differentiating factors become, instead, productivity and price. Incidentally, these elements are already at the core of many LSPs and platforms’ job allocation strategies, underscoring a perverted practice: that the system encourages suppliers to bid lower than is sustainable for securing endless pieces of fragmented work. It seems that language professionals, in a Squid Game fashion, will have to fight a tug of war.


Commodification and the threat of interchangeability

Language professionals also confront complex perceptions of work value. The knowledge economy includes forms of economic activity that place ‘immaterial’ knowledge at its heart. In fact, language work is an example of Hardt and Negri’s ‘immaterial labour’, the production of intangible goods such as communication delivered by temporary productive units of people. The intangible aspects of language services make it difficult to identify and standardise both their market value, but also output quality in a repeatable way that aligns with user and corporate expectations. Take the Squid Game controversy. Audiences lamented the quality of the translation, stating that important aspects were not preserved in the closed caption, subtitling and dubbing from Korean into languages such as French, Hindi and English. However, many translation studies scholars say that the criticism was unfair, as spectators do not fully understand the intricacies of language work. For instance, no matter how long or complicated the original dialogues, it is complex to condense messages and culture-specific elements into one-inch tall slots, or to convey ‘untraslatable’ concepts which do not have an equivalent into another language. Additionally, lack of understanding and of clear market and quality value create the notion that language work is a commodity like many others. According to Chris Fetner, the director of the Entertainment Globalization Association (EGA) and a past Netflix executive, outputs such as subtitles can be “very good, but when you try to go to perfection, the return on investment becomes uninteresting. Having something that’s 95% satisfying […] And to move that 5%, it’s expensive.” Language professionals took to social media, responding that the return on investment remains interesting because industry rates are low, and that contractors should pay professionals fairly instead of treating their work like a meaningless commodity. The implication is that language work is not perceived as a source of value creation but as a production factor which bears costs but offers limited competitive advantage.


The futures of language workers

Automation and platformisation of labour have subsumed a vast part of the language industry under the rule of techno-capitalism. As for many other sectors, the organisation of work in this industry has not become more direct, democratic, or valued. Instead, it has led to the development of exploitative practices and the low “signalling” of language professionals’ labour value. What this amounts to, particularly, is a lack of visibility—the preference for automated solutions rather than more costly, slow human intervention which obscure the role and skills of professionals. These dynamics point to a troublesome relationship that fuels mechanisms of devaluation and dehumanisation, where the language worker becomes yet another invisible cog in a largely automated process that directs the purchasing and selling of language as a commodity.


Deborah Giustini is a Postdoctoral Fellow at KU Leuven.


Image credit: Andrés Rodríguez from Pixabay