“The best way to predict
the future is to create it”
Peter F. Drucker
(Father of Modern Management)

The translation business model, as it has evolved since the dawn of history, is no more—but only to a certain point. A new era has begun for both translators and interpreters, albeit at different levels and to varying degrees. Neither one will disappear any time soon, if ever. Agency, whether human or machine, or artificial intelligence (AI) in particular, has well-known limitations. Explaining these and recognizing their real overall impact helps to provide a clearer picture of how institutional translation might develop over the next decade. To deliver fully on their mandate, management and translators need to take a more critical view of today’s nascent technologies, fully leverage their promised “added value”, and remain vigilant about the direction in which big data and human generated knowledge are heading.

Limitation One: Translation and Artificial Intelligence’s (Pandora) Black Box

Understanding AI-based machine translation is understanding first AI limitations. A quick review of specialized articles shows that AI is gradually turning into an unbound geopolitical race. Proponents and opponents hold utterly opposite views over its “boundless” promises or “existential” threats. World powers are rushing to test its potential in military manoeuvres, logistics, maintenance and even in combat zones. Relaying specialists’ opinions, MIT Technology Review, the Harvard Business Review, and the Economist magazine speak in this respect of “Spota”, “Maven”, and “Lavender” projects capable of speeding up operational military decisions million folds. Increasingly, human crucial decisions rely on AI output and analysis. Human agency is slowly taking the back seat in some niche fields. Yet, taking a close look beyond the ongoing positive narrative, understandable from firms’ commercial vantage point, AI has a fundamental epistemological flaw, nonetheless. Until today, it remains a black box in that nobody, not even its developers, know how it reaches this or that conclusion and by what means or method. There has been some timid attempts to decipher its workings based on its output. The results are still lacking. Worse, AI is far from standing the test of Popperian falsification principle, and no serious scientist could afford to ignore this. It is no surprise then that the European Union introduced this August a new Law regulating its use, in a push to fend off the potential and/or known risks involved. Some prominent insiders and political leaders, including UNSG Mr. Guterres, describe it as an “existential threat”

Limitation Two: Translation and Human Cognitive (Over) Load

Understanding the translation process is appreciating the cognitive (over)load and its boundaries. Studies conducted by John Sweller in the late 1980s for instance show that human cognition is an “intricate and highly developed system”. However, it has inherent limitations which impact significantly human ability to process information effectively. One of the critical concepts in mapping these limitations is cognitive load or the “amount of mental effort spent by the working memory”. Cognitive load theory provides thus insights into how the load can affect learning and performance. It is divided into three categories: intrinsic, extraneous, and germane. The first is irrelevant as translators already master the source and target languages. The extraneous refers to the additional cognitive effort imposed by how information is presented. Poorly drafted texts do increase the extraneous load, making it harder to absorb the intended meaning, nuance, ambiguity and context. Germane load, on the other hand, is the mental effort devoted to processing, constructing, and automating schemas. The implications of cognitive load are therefore profound. When it exceeds the capacity of working memory, understanding and performance suffer. Translators are no less prone to such a limit, and shouldn’t be expected to do what is beyond their real natural capacities. Technologies do help fix tedious routines, but shouldn’t be considered “Holy Water”.

Limitation Three: Hybrid Translation or the Best of “Two Worlds”

Human-machine hybrid translation calls for combining the strengths of human and machine to overcome their respective limitations. When specifically applied to business-related texts, MT is fast and can be cost-effective. When applied in highly specialized domains, such as international relations, negotiations, and law, it tends to mishandle ambiguity and terminology, lacks creativity, and blatantly misses cultural nuanced information. In contrast, human translation provides accuracy and nuance but is time-consuming and somewhat “expensive” at times. The hybrid approach favours then having machine generating initial drafts, and humans refining them, allowing thus faster, more consistent, culturally appropriate and fit-for-purpose texts. Nonetheless, a closer look shows that language pairs are not born equal, especially minority languages where the data is scarce. Therefore, this optimism has yet to be borne out across the board.

It is a fact that a gifted translator is unlikely to be an equally gifted reviser. The reverse is no less true. The two tasks and skills are totally different. Therefore, seeking to morph a translator/reviser into a post-editor, based on sporadic training and over a short period of time, is no less a fallacy. The premise has always been that a translator is a writer first, then a translator who, like wine, would mature on the long-term only. To become a post-editor of MT unmediated texts would require a much longer path and a steeper learning curve. There are no technological shortcuts to a genuinely reliable full-fledged professional. Translators require considerable nurture and this takes time. One simply can’t translate what one doesn’t fully understand, no matter the technological tools given. Believing that the likes of MT/ChatGPT, churning out unchecked and unvetted information, would do the trick is short-sighted and naïve at best. Translation has never been the sum of tokens, algorithmic calculations and syntagmemes. To sum it all: Machine Fluency is not Human Accuracy.

Future Trends, Gatekeeping and Concrete Risks

Existential Threat means generally hollowing out human generated knowledge. It is a fact that very few people today know how to grow food. Maybe tomorrow translators wouldn’t be able to work on their own, should and when the need arises. So, hurrying to shift translation processes and agency to MT at a time when the dust hasn’t settled yet on its real potential carries certain risks for all stakeholders, primarily management. Some well-established US firms have already started experimenting with avatar CEOs. More still, high-quality MT systems development and maintenance require intensive resources and raise at the same time serious security and privacy concerns in the absence of tough legislations. Masked Language Modeling (MLM) training is said to be on the cusp of running out of freely available human generated data, public-domain (UN) institutional data included. Working on a new cycle of “synthetic/fake data” to train MLM models has just begun as copyright owners are fighting back against unauthorized use of their material.

As for UN institutional translation per se, the near future will depend on at least three major factors: Budget allocation, degree of management’s (and somehow member states’) embrace of new technologies, and translators’ appreciation of their indispensable and crucial role. The first two are somewhat outside translators’ immediate control, but they can still leverage and positively influence translation final outcome. For reasons long known and researched, translators/interpreters will remain the ultimate gatekeepers and guarantors of cross-cultural politically nuanced communication. Text ambiguity is one course purposely taken at times by parties and interlocutors to circumvent sensitive issues. Institutional texts are awash with these intricate paragraphs.

Human translation is old wine and MT is another new bottle in that tools are changing, but the semantics of translation are not and … will not. Translators, and interpreters no less, need to recognize their pivotal role and immense capacity to create, in Drucker’s words, their own future, and shape it for generations to come. It is in management’s interest to support them as their mutual futures are intrinsically intertwined.

If you are able to access the United Nations sharepoint, you can find the link to the AI augmented version of the article here.


READ MORE ARTICLES FROM 

INSIDE VIEW