In-house translation – outside the box

Category: Terminology

  • Does shadow translation present a bigger threat to language services than MT/GenAI itself?

    Does shadow translation present a bigger threat to language services than MT/GenAI itself?

    As MT and GenAI offer seemingly free and instant translation, “shadow translation” is a reality for in-house language services. There is an inevitability in corporate life, that prohibiting people from taking shortcuts, will not stop them from doing so. Similarly, even if there is a path, there will always be a desire path. Well-meaning rules, e.g. banning free MT tools will always be broken, unless sites or tools are blocked.

    Such practices are typical examples of shadow translation activities. In organisations with a team of in-house linguists, shadow translation activities often form due to the bottlenecks in translation via language services, i.e. the need for translation outside the official, governed translation process.

    Shadow translation serves as an unofficial, parallel or informal translation conduit. They are produced outside of authorised translation workflows. Shadow translations also evade capture in translation management (e.g. job lists, translation memories, or terminology systems). However, they remain in use in the wild by staff members, from trainees through to board members.

    What kinds of translations are vulnerable?

    But what kind of translations do shadow translation cover? Frequently it may be translations lazily labelled as “for internal use” i.e. for quickly getting into the meaning of a text, for writing something up etc. But, it might be used by staff, if someone doesn’t feel confident enough to draft texts (e.g. e-mails, summaries or speaking notes) by themself. In some situations, staff members may even delegate translation to the non-professional translator bilingual colleagues. This is particularly visible in job descriptions for PAs with an expectation of multilingual written communication.

    In the MT/GenAI era, there is the complicit use of MT/LLM-based translation, often without subsequent review (i.e. the translation, due possibly to its deceptive fluency, is passed on “as is”) or not disclosed as being a translation. However, it may include cases of externally translated content that sits outside of internal translation storage systems.

    This can also result in siloing in departments and divisions, meaning that there are multiple diverging translations in circulation of the same texts (for example in the case of standard letters, text blocks etc.) This is one reason why I translate each iteration of the standard texts that form administrative decisions for the entire organisation. I ensure that there is a single source of truth, and then only need to focus on the ruling (Spruch) of the individual administrative decision in hand. In process management, where processes contain sample texts about decisions and mails, I also make sure only to include a sample ruling. Other parts of such sample texts refer to the current template to use for the rest.

    Is this an issue for translators?

    Shadow translation bears numerous risks. From the perspective of a language services unit or a translator in one, the sensitivity is different to that experienced by the users. Users are likely to have a slightly more cavalier approach, and see it as a “victimless crime”.

    Unofficial translations, without quality assurance processes, may diverge subtly from the legally binding or authentic wording of a translation. To provide an example, I recently saw a Copilot Agent use case. The purpose of the agent was “legal translation”. The agent saved parallel texts of a piece of EU legislation in two languages as “knowledge” into the agent. The prompt in the agent instructed it to translate a specific provision from one language into another. Without the prompt stating the Spanish text was an authentic translation of the English text, the agent decided to create a back-translation for its English version.

    This can quickly create issues (e.g. like distorted messages in a game of Telephone / Stille Post). It leads to parallel translations that are close but not identical. Potential outcomes can be incorrect wording invalidating decisions or misrepresenting obligations. This results in exposing an institution to being challenged about their decisions.

    But it’s “only” for internal use, right?

    The problem with claiming that something is “for internal use” is a downstream one later down the line. Months or years later an internal text may suddenly become an external one. When this happens, it is like a document crossing an institutional firewall. Translation in language services can frequently be a “black box” for non-translators. Translators on occasion lose themselves in their own processes or their perceived reality. In the worst case, they might actually neglect that they are translating for a specific audience (e.g. the “Clientland” analogy that Chris Durban frequently alludes to). Translators need to also consider how they label translations – and how it will be (continue to be) used after they deliver it to the customer in-house.

    Another issue from shadow translation is that it supports terminology drift – with key terms gradually changing their meaning across documents, or where different departments resort to different equivalents for the same source term, or where informal translations take on a reference quality. Variants can slowly leach into accepted use, and are difficult to reverse. In the human-machine translation age, the effects may be more pronounced, as the training of models has shown, as once something is in a model, it can be very difficult to get it back out of a model.

    A question of accountability

    A further issue for language services can be down to the emergence of accountability gaps. The assumption might be that the error occurred in language services. Within language services, translation memory systems can resolve the team members involved. In turn, it will also identify the customer, reviewer and who signed off on the translator. Shadow translation can’t do that, other than identifying the string of communications. Nevertheless, language services may be caught in the crossfire. It can have the ability to undermine the authority of the official translation function in the blame game. Inconsistencies that they never contributed to nevertheless land at their door! Invisible translation output can also render quality metrics meaningless.

    Is shadow translation an accidental or malicious phenomenon?

    Ultimately, structural pressures are usually behind the creating of a burgeoning level of shadow translation. These pressures are due to time pressure (everything required now, or better still yesterday!) Available resources in language services might be creating the bottleneck – there are not endless supplies of human translators in all language combinations. AI tools seep out of every pore in corporate life (especially due to the “fear of missing out”!) unwittingly casting translation as being readily available on tap. Another phenomenon has been that of “multilingualism = translator” conflating speaking more than one language to being able to translate. Even children raised as bilingual/multilingual from birth will still gravitate towards one language over another in specific circumstances.

    However, sometimes the issue is down having a lack of clear rules, in order to distinguish working, informational, or legally or externally relevant translations. These issues are scaling silently as digital tools develop and potentially even improve.

    So this isn’t just about a translators trying to protect their jobs?

    When translators try to explain their continuing role in the GenAI-dominated landscape, they are accused of being protectionist. Persona prompts apparently can make everyone an “expert translator specialising in xyz in the language combination A to B”. But the institutional impact is more about governance rather than an out-and-out language one.

    Shadow translation circumvents document control, risk management, process transparency and accountability – a point missed by trivialising it as just being a discussion about quality. In this regard – it ties in a lot with the shattering of the triple constraint for translation, that I refer to in my blog posts about the “Expert in the Lead” approach. Particularly, given the fact that human in the loop approaches in translation do not adequately consider risk management and accountability.

    The “internal use only” caveat also is a risky approach. Internal documents (e.g. briefing notes) actively shape decisions, which in turn shape actions, which themselves have external and legal consequences. And internal translations can end up becoming de facto reference documents.

    Where there is an added dimension due to AI and MT is due to the fact that such tools also play a role in making shadow translation faster, more tempting to use, and harder to detect. Prompting courses invariably mention using prompts to try to “make the text not look like it is AI-written”. This is why institutions need clear policies to avoid uncontrolled MT output, inconsistent style and terminology, as well as false confidence and over-reliance on “good enough” MT output.

    We need to talk about cognitive load and cognitive downstreaming

    I increasing mention the concern about cognitive downstreaming. Offloading (linguistic) effort to tools or other individuals downstreams review, validation and responsibility involved. If you use a linguist only for the review stage, they have to try to make “informed” decisions without the full information – I’ve started to receive “Copilot translations” (the quotations are intentional!) in English – without being provided a source text. I enquire where the original German text is. After all, if it is a translation, it must be a translation of a German text.

    Sometimes it even emerges that the text was in fact prompting output. However, even the detail in the prompt provides me with information that is missing from the English, e.g. about the legal provisions forming the underlying basis (and quickly highlighting that there has been an unnecessary back translation). In the worse cases, the output has read like LLM output, where the German original also doesn’t respect the house authoring and formatting style.

    What should institutional practices do to get a grip on shadow translation?

    There is a clear near for greater formal distinction in the classification of translations as an approach to shadow translation management. Labelling is an issue that is emerging with the automation of translation of web content. I am fully aware that bad human translations also exist! Take a look at many multilingual tourism websites for private accommodation if you want!

    One potential categorisation could be as follows:

    • authoritative translations (binding, publishable)
    • working translations (informational, internal)
    • machine-assisted drafts (explicitly non-authoritative)

    Within each category, the permitted uses, visibility requirements and review expectations also need to be specified.

    Let there be light! Visibility of unofficial translations

    The way that shadow translation has shown that banning a tool only means that someone will look for the next tool to circumvent the ban, is possibly a sign than a more enlightened approach is to not ban shadow translation (as a process or its output).

    I’d suggest destigmatisation by recommending to improve their visibility. However, there would be some requirements, namely:

    • enforced use of labelling (e.g. also involving the categorisations mentioned above).
    • requiring storage in a shared environment, with the source text available in the same location.
    • compulsory disclosure about the use of MT/GenAI (this can also help reduce cognitive downstreaming).

    Favouring a visibility approach, over prohibiting shadow translation, can also help ensure a steady work for professional translators. For example, quality scoring of shadow translations can form a ranking for permit the prioritising items for translation to ensure greater availability of authoritative translations.

    Visibility on the other hand, can also mean improved centralisation and availability of language data (e.g. from translation memories or terminology databases). Two such approaches are what RWS calls “generative translation” or Kaleidoscope calls “TAG (terminology augmented generation). I’m hoping to be able to convince my employer to allow more effective terminology sharing than in the old-fashioned “bilingual glossary”. That product is no longer cutting edge, and is high maintenance. Each update consumes a lot of time for a product in an old-fashioned format.

    The role of translators in AI governance

    Back in 2022, I joined our Innovation Lab, and chose from the outset to take a “voice of reason” approach, which would question use cases, AI governance issues and the like. I tried to be open to innovation, while also ensuring that the recognition of the continuing importance of human translation . They recognised and understood my valid concerns. I emphasised that I was risk aware and possibly somewhat risk averse. Labelling myself the “Spaßbremse” (spoiler of the fun) prevented anyone else from using it in an accusatory tone.

    I contributed towards my institution’s AI governance – by highlighting the need for review, and highlighting issues of accountability, reputational loss, and credibility. More importantly, it guaranteed my seat at the table. Langtech solution providers currently struggle to be heard. BigTechs demonstrate their (lack of) knowledge of translation and seem hell-bent on ignoring incumbent niche players. In this regard, my approach was about being heard rather than just to be an end-user.

    I’ve also used it to cultivate my role as a “risk sensor”. Deceptive fluency of MT output frequently fails to identify inconsistencies in source texts, ambiguities in policy, legal uncertainty and procedural gaps. Trying to bring the issue of “shadow translation” onto the table is also a case of highlighting that it bypasses governance, rather than my being protectionist or territorial.

    So what about a bite-sized takeaway?

    Shadow translation is a hidden reality of any organisation with a multilingual environment. Failure to govern shadow translation will lead it to self-govern. Self-governance is inconsistent, invisible and allows the propagation and permeation of risks.

  • 7 thoughts on the challenges facing in-house translators

    7 thoughts on the challenges facing in-house translators

    At the start of this week, I attended an event organised for in-house translators by Universitas. I was part of a small and eclectic group, meeting at the end of the working day. From a very interesting and relaxed couple of hours talking with the others in the group a number of things sprung to mine, which I addressed the following day on LinkedIn.

    1. There is no “one-size-fits-all” in-house role. Expectations on in-house translators vary dramatically: from subject matter specialists through to multi-language pair generalists. There is no hard and fast rule about work in a single language pair or multiple pairs. Even classical “Translator” positions may involve a mixture of translation and other activities. For applicants for such positions, it is advisable at interview to ask about the likely expected hours dedicated to translation.
    2. Job descriptions for open vacancies are seldom as straightforward as “Translator (m/f/d)”. Translation is frequently only a part of the job description, and the job title is seldom for only a translator. In language career portals in the German-speaking world (e.g. Stepstone, Kununu or Karriere.at) most hits for the skill “Übersetzung” talk about it figuratively.
      For example a recent job advert for Österreichische Post AG for an Expert in Controlling Insights mentioned “Du fungierst als “Dolmetscher” zwischen Fachbereich und Programmierer und bist zuständig für die Übersetzung der Anforderungen des Fachbereichs in detaillierte Vorgaben und Zielsetzungen für technische Umsetzung.” In other words – nothing to do with translation or interpreting! In other cases it might be disguised in a job description for a “zweisprachige Schreibkraft” or “Kommunikationstalent”. This latter role is closer to transcreation than translation. Here are some common language-based job titles.
    3. The burgeoning TechStack: among the group around the table, the tools used, and expectations regarding such tool use was varied. There were also varying views about the expectations regarding the use of GenAI/LLMs. Some are very open to the possibilities, while others actively decline to use such tools, or are not permitted to do so.
      Often required training regarding GenAI/LLMs is not specifically tailored to translators. Similarly, the effective use of such tools for translators is not as clear cut as the hype makes out. Deployment of new tools is often IT-led. This approach sometimes overlooks those with genuine expertise to really use them and assess the quality of their output.

    Resource issues

    1. Double-hatting is common: some translators also work as interpreters, rather than having separate translation and interpreting personnel. Others in-housers are only part-time in-house, so have to juggle self-employment alongside their fixed employment.
      This naturally places additional demands on them in terms of time management and also how to organise their time effectively. Teleworking may have cut out some unnecessary miles/kilometers, but there is still a lot of juggling required with multiple positions.
    2. Working for a demanding customer base with dwindling human translation capacity: this can become even more difficult if language services are overseen by non-linguists. This can make it more difficult to discuss the need for quality that goes beyond “good enough”, and where “fit for purpose” is a minimum requirement.
      It can be difficult to get past only being viewed as a cost centre. Translation can be quantified easily in terms of cost, but its impact on sales etc. is more difficult to quantify. Tighter budgets mean fewer retiring colleagues are replaced, or FTEs are replaced by fractional headcount. Alternatively FTEs might only have a certain percentage of their time devoted to translation.
    3. Decreased job security: even in public administration there are perceptions that the job security of translators is lower than it used to be. The erosion of the classical triple constraint, the rise of “good enough”, and the improved “linguistic fitness” of many white collar colleagues has affected demand for translators.
      Job mobility and exchange programmes while studying mean that many colleagues are more confident in their language abilities that only a few years ago. However, there is still a subjective basis to their assessment of their own language abilities. Just as having two hands and a piano does not make me a concert pianist, working knowledge of two languages, does not automatically transfer into being able to write well in your target language.
    4. Rising expectations in terms of output: while tools like CAT and (N)MT have helped to increase translator productivity, there is still the unrealistic expectation in light of the promises of “instant translation” offered by browser-based tools.
      Translators’ potential output can really depend on so many factors – NMT/GenAI/LLMs are “confidently wrong” – they will always offer a translation, whereas the “cautiously correct” human translator reverts to the author if unsure – to clear up potential source or target text ambiguities.
      Similarly, expectations vary wildly based on the percentage of time spent on translation compared to non-translation activities. Often there is no dedicated capacity for terminology work. Only larger language units have dedicated terminologists: without them, it is often widely neglected. With the advent of MT/GenAI and the Terminology Augmented Generation approach, which is used to import your terminology into the LLM, it is likely to gain in importance.

    Are you interested in events like this? Universitas holds regular events throughout the year. Check out the Universitas website for more information – if you are not yet a member, some events are open to guests. If you are interested in knowing more about what I do, then why not join the Universitas Berufsbilder webinar on 23 October 2025, which will focus on the role of project management and process management.

  • Even Homer Nods! How Chinese Walls showed it was good to talk

    Even Homer Nods! How Chinese Walls showed it was good to talk

    One downside to being the only linguist in your setting and with the knock-on of limited exposure to your target language is a lack of opportunities to discuss intricacies of usage. This issue can of course lead to an inadvertent usage, or use of a non-inclusive term. The scenario was this: I was engaging on a post on LinkedIn about why it is good for translators to be able to talk to one another in projects where multiple translators work on the project.

    The isolation as a translator means that you can often become blinkered in your vision (and thinking!). As wonderful as stable customers are – and they are increasingly rare in the cut-throat “first-finger-first” world of getting translation work. Established translators have interesting ideas: translation slams, retreats, networking events and the like. All aimed at talking. Collaboration can be a very rewarding experience – this is why I appreciate the opportunity of reviewing EBA Guidelines – I get to ask questions, look at the translation through a different prism or lens.

    Regarding collaboration and communication in the post in question, my well-intentioned comment was:

    […] collaboration thrives on direct communication between translators – sometimes agencies throw up unnecessary Chinese walls between translators working on the same project (I remember once working on a large project where we were all only allowed to “chat” through a moderated anonymous portal – such was the fear of the agency that we might build a team that would compete against the agency we were working for!) In another crazy situation years ago, where a large project was carved up between two agencies, I discovered that I was providing terminology support for the same project I was translating part of for the other agency.

    Bang. And then it was. A careless choice of words, although one I knew from a professional context.

    Even Homer Nods…

    The poster responded by pointing out that “Chinese walls” was an outdated and offensive term for some. I chose to own my mistake, accepted their insight and replied that I would refrain from using it in the future. A hollow promise? No. I immediately acted on it. How did I act?

    1. I immediately accepted the mistake – and said that I would refrain for using it in the future. With reflection, and hindsight too, I realise that the analogy is also incorrect/lazy – both in terms of my usage and in a financial context. The Great Wall of China (and walled cities throughout history) served for protecting territory, but not to prevent citizens from leaving. Suitably chastened, the next goal was to find a more accurate term for the required context.
    2. I asked the original poster whether they can suggest “a suitable, succinct and more inclusive replacement for its use in the sense of “a barrier to avoid potential conflicts of interest” e.g. between advisory and trading divisions (as used in investment banking)?” The OP promptly offered terms like “firewall” and “screen”. My own research also yielded “ethics walls” and “ethical walls” – as mentioned by Judge Low in Peat, Marwick, Mitchell & Co. v. Superior Court (1988). Low’s suggestion also works particularly well for preserving the concept of “hiring over the wall”. This is an internal staffing move of hiring someone from the other side of the ethics wall.
    3. I amended the terminology entry for Chinese wall(s) in my MultiTerm termbases. It now appears as non-inclusive (an additional field at term level used in conjunction with a QA tool for inclusive language). This means that in the event of any German term translated as “Chinese wall(s)”, I receive a warning that the term is to be avoided, to allow me to then correct it as I go. This is particularly useful in terms of fuzzy matches. In addition I have a MultiTerm export routine that allows an export of non-inclusive terms. Its purpose is to increase awareness by circulating this list regularly, to help non-L1-colleagues.
    4. Checking existing usage in publications. For the website, I ran a google search for site:fma.gv.at “Chinese wall”. Hits appeared mainly in downloads from external sources, or from historic documents – on both the German and English versions of the site. These I can’t change, because they are the words of other stakeholders. When we used SiteImprove, I used reports (e.g. about the use of Chinese *AND* wall). Using reports in SiteImprove frequently identified obsolete terminology for correcting it. I contended with a lot of “inherited” translations from numerous translation sources, prior to becoming our English web editor. I discussed any mentions found in the German version with post/page authors.
    5. Checking existing usage in translation memories. I check both the source and target texts of segments for usage of “Chinese wall(s)” and how it was translated. For example, the English source spoke of “Chinese walls” but the German target spoke of “Informationssperren (sog. “Chinese walls”)” In some cases, I remove the translation unit from the Translation Memory. In other cases I add a quality penalty (I use a numerical quality field from aligned texts that I used to build up my initial translation memories, based on their quality (e.g. 95-100 for aligned texts of legal acts from Eur-Lex (with penalties for age), 85-90 for aligned texts of publications where the emphasis of the translation was not such a strict word-for-word rendering, (with penalties for age)). I delete translation units “past their sell by date”. (I’ll deal with TM housekeeping in a future post).

    However, I didn’t beat myself up over it. Instead, I used it as a way to test my processes and workflows, which work well. I also used it as a way to think of several initiatives for informal networking events to could draw positively from the experience. And I took it on board as a shortcoming of living outside of my L1-target language culture. All of which show the importance of collaboration and talking.

    It’s Good to Talk…

    Reverting to the original post that I commented on, I remarked about my first physical meeting of my Working Group in Athens after the pandemic that “It’s good to talk!” In this regard as a lone in-house existence is similar to that of many freelancers in that it is a relatively isolated one. However, only if I choose it to be. This is why I try to exploit opportunities to talk to my colleagues in my department about their policy areas. This is how I remain alert to new policy information, legal acts, soft law instruments, issues encountered in operative supervision. These are important for my improved understanding of operative banking supervision.

    Some pearls of wisdom yield new aligned texts (e.g. new guidance on aspects of banking supervision). Part of my non-translation remit is also keeping a watchful eye over our banking supervision processes. This is a quality assurance and quality management process, and provides useful insights for translation purposes about legal developments. Translators talking and collaborating helps them to discover new opportunities, gather different perspectives and bounce different approaches off one another.

    When I outsource a piece to someone, I believe they should receive as much information as necessary. I believe in being approachable where they have questions and to discuss terminology issues. Otherwise I ensure that they can directly access the author of the piece to translate. After all the author knows best what they mean to say. In my response about translators being able to communicate, even someone pointing out my error confirms the value of communication.

    The final word…

    Dialogue is a two-way street – the more you give/contribute, the more your receive. In addition, this episode resulted in connecting with the OP, whose content I have read for a number of months, coming up with a creative idea for the future, and a blog post.