7 thoughts on how the Human in the Loop approach trivialises the translation profession

The Human in the LoopHuman in the loop is an approach advocated by the translation industry, whereby machine translation is used and the role of human translators is secondary – e.g. doing some kind of post-editing. It has become particularly prevalent with the rise of LLMs. My issues with HITL is that it reduces the role of a human (with no mention of their More approach trivialises the translation profession in many ways. Here are just seven of them, based on some recent discussions on LinkedIn in various threads of comments on posts.

1. It fails to set out the qualification/expertise required to be the human in the loopHuman in the loop is an approach advocated by the translation industry, whereby machine translation is used and the role of human translators is secondary – e.g. doing some kind of post-editing. It has become particularly prevalent with the rise of LLMs. My issues with HITL is that it reduces the role of a human (with no mention of their More responsible for ensuring the quality of the output. The human may not necessarily possess the genuine level of language skills in the target language to assess the quality of MT output. In particular, non-translators should not be allocated the task of doing the heavy lifting for e.g. appraising the quality of an MT solution.
2. Advocating a “good enough” approach that is not “fit for purpose” nurtures the narrative that anyone with a better than rudimentary grounding in two languages is a translator. Translators frequently given the inner eye-roll upon hearing qualification of source/target language ability via statements like, “My brother’s son grew up in the States” or “I spent 18 months in Brussels”. In both cases there is an exposure to the language, but there is no guarantee that they have any formal training or education as a translator.
3. By relegating the human in the loopHuman in the loop is an approach advocated by the translation industry, whereby machine translation is used and the role of human translators is secondary – e.g. doing some kind of post-editing. It has become particularly prevalent with the rise of LLMs. My issues with HITL is that it reduces the role of a human (with no mention of their More to a role that is often no better than lightly edited MTPE for derisory rates, it means practitioners may only scrape a minimum hourly wage. Translators on minimum wage will as a rule not be as committed or engaged as if they were being remunerated commensurate to their skill. Similarly, this approach is also a classic example of how translation has been uberised to the extent of reducing translation to a commoditised service.
4. The real cost of GenAI translation is questionable: the amount of energy consumed on LLMA large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content. More training and to use the LLMA large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content. More to “translate” is not apparent – skewing the real cost-saving. If the output is of such poor quality that it has to be (re)translated from scratch, the cost is high than using human translation in the first place. Where models use multiple alternatives for single sentences e.g. where there are QAQuality assurance (QA) tools in Trados can be used for example to check that the terminology used corresponds to that in a Termbase. More steps also performed with an LLMA large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content. More, the number of prompts used may be vastly more than one per sentence. The metric of comparing the among used for a search machine request cf. a prompt is no as reliable as it was, as many search providers now offer AI summaries automatically, driving up the energy consumption of a search request.
5. Linguistic colonialism and extensive use of pivot languages unrealistically trivialises LOTELanguages other than English is used to describe any source-target language combination that does not involve English. It is abbreviated to LOTE. More (Languages Other Than EnglishLanguages other than English is used to describe any source-target language combination that does not involve English. It is abbreviated to LOTE. More) (h/t Sarah Swift). Frequently models are trained between the source language and English, but less frequently directly between to LOTEs. Translation using English as a pivot languageA language used in machine translation as an intermediate step in the translation from the source language (SL) to the target language (TL) in the event that there is no bilingual engine directly between SL and TL. For example there might not be a bilingual Maltese to Finnish model, as so English is used as the pivot language. I.e. MT More may overlook similarities/proximities, or even the abundant supply of translators in a language pair between geographically neighbouring languages (e.g. consider translation directly from Slovak to Hungarian, as opposed to an MT route using a pivot languageA language used in machine translation as an intermediate step in the translation from the source language (SL) to the target language (TL) in the event that there is no bilingual engine directly between SL and TL. For example there might not be a bilingual Maltese to Finnish model, as so English is used as the pivot language. I.e. MT More, where a Slavic language to a Finno-Ugric language is performed via a Romance language pivot). Pivoting via English introduces unnecessary lexical gaps.
6. MT’s “confidently wrong” approach is at distinct odds with a professional translator’s “cautiously right” approach. The latter will use context and question wordings – and refer the source text back to the author, whereas the former will always offer a translation. If you want to test this, take a provision from a law with nested clauses and remove parts of the verb phrases. The MT tool will nevertheless always offer a translation (based on probability) without context, whereas a human can at best guess from the surrounding context, or state that there is a problem with the source text.
7. There is even a risk that MT might create a false veneer that translation is a job that “anyone can do” – thereby reinforcing the false impression that translation is little more than a mechanical task of replacing words, rather than a expertise-based service delivered by highly specialised trained professionals.