AI translation – an opponent’s perspective

On March 30, 2023, Mochiwa Mochiya Pty Ltd (“MMP”) published an article on LinkedIn titled “AI Translation – A Translator’s Perspective.” It is at once thoroughly serious and disarming in its tone and strategy. On the one hand, MMP tries to maintain an air of humor and informality. On the other hand, MMP seems to have taken seriously Alexander Pope’s barbed advice to design a piece of writing “like a labyrinth, out of which nobody can get you clear” but the writers themselves.[1]

Still, MMP seems to have left clues. And it is not unfair to say that whatever else MMP intends, it intends to persuade itself and everyone else that opponents of AI translation (DeepL, for instance) are “utterly bad and utterly stupid.”[2] That they are either derelict in the performance of their duty or doing the industry a disservice or both. MMP writes:

Last time I checked, translators have a duty to provide the best possible translations to clients. If you’re not utilising the latest AI technology to aid you in this endeavour, you’re quite frankly, doing the industry a disservice.

Now, if MMP had arrived at this conclusion via careful argument, opponents of AI translation such as myself would be obliged to see and feel its point. But MMP arrives at this conclusion via a slew of rhetorical no-no’s. Here’s MMP in its own words:

In case you didn’t realise it by now, we are in the early stages of the digital revolution where keywords like AI, blockchain technology, cryptocurrencies, and other web3 technologies will take centre stage in all aspects of our lives. Yes, you can resist such developments, but the truth is, refusing to embrace AI technology in translation will only leave you lagging far behind your more forward-thinking counterparts. I mean, what is the basis of such arguments against AI? Last time I checked, translators have a duty to provide the best possible translations to clients. If you’re not utilising the latest AI technology to aid you in this endeavour, you’re quite frankly, doing the industry a disservice.

There’s a lot to unpack here. So, let’s start at the beginning. If MMP had avoided indignant language up until this point, which it doesn’t, it might have gotten its view of the facts straight. As things stand, however, the black-letter interpretation of its words is that MMP confuses the digital revolution with a vocabulary list, which has yet to take center stage in all aspects of our lives. While the digital revolution has brought with it new vocabulary, the digital revolution does not consist of that vocabulary. It consists of new digital technologies.

But let us grant MMP a straight view of the facts, that the digital revolution is not where keywords, but where digital technologies like AI, blockchain technology, cryptocurrencies, and other web3 technologies will take center stage in all aspects of our lives. Even this grant of facts fails to entitle MMP to the conclusion it wishes to draw here. For MMP sets us a paradox; it ignores crucial circumstances, which it itself supplies. In its own view, MMP grants that these digital technologies have not yet taken center stage in all aspects of our lives; its view is that these technologies will take center stage, which is to say: only at some indeterminate point in the future will these technologies take that stage. But if they have not yet experienced widespread adoption (I take the phrase “center stage in all aspects of our lives” to mean ‘widespread adoption’), the question arises: how could anyone be resisting or failing to embrace technologies whose widespread adoption, in MMP’s own view, lies only in the future?

Perhaps seeing a paradox here is doing MMP an injustice. Perhaps MMP’s view is not a paradox at all. Perhaps MMP wishes only to draw an analogy between the translators in “staunch opposition” to AI translation and “flip-phone enthusiasts who once claimed that the iPhone would never catch on.” MMP seems to believe as much, and by extension, it seems to believe three things, which are not simply untrue, they are ignorantly and comically and ironically untrue.

First, MMP suggests that flip phone enthusiasts were left “lagging far behind [their] more forward-thinking counterparts [= early iPhone adopters].” What? I lived in New York City when the first iPhone launched in 2007 (I even had a flip phone). Only the most hardcore Apple fans and people who could stomach AT&T’s then-horrible phone service were early iPhone adopters.[3] Everyone else, myself included, wanted to wait, and ultimately did wait, until more information on the technology itself and additional carriers became available. Besides the lack of information and carrier options, there existed a question as to why anyone would want internet on a cell phone; in 2007, that was a brand new idea, and people had to warm up to it and come to understand its benefits. And nobody thought they were lagging behind early iPhone adopters.

Second, MMP suggests that, like translators today, flip phone enthusiasts had a duty to provide the best telephone calls to their clients and were doing the telephone industry (or Apple?) a disservice by not utilizing the latest telephone technology. Why should flip phone enthusiasts have failed to provide the best telephone calls with a flip phone? Even a Donald J. Trump would have to know this is false; the former guy thinks he has the best, nay, the most perfect phone calls with landlines. And why should flip phone enthusiasts have some duty not to do the telephone industry (or Apple?) a disservice? Were all flip phone enthusiasts employed in that industry (by Apple?)?

Third, MMP poses a rhetorical question – “I mean, what is the basis of such arguments against AI?” – and suggests that, like translators today, flip phone enthusiasts had no good reason not to be early technology adopters. The irony here is that MMP knows that translators today have good reasons not to be early adopters of AI translation tools. MMP knows this, because MMP itself provides three good reasons of many.

First, despite AI translation tools’ improvements over earlier rule-based and statistical machine-translation tools, “AI translation still makes mistakes, struggles with certain nuances, idioms, highly technical [= specialized] content, and high-context culture.” In other words, AI translation tools struggle with everything that makes a translation a translation. Second, AI translation tools can harbor “biased algorithms or datasets” that

may perpetuate harmful stereotypes or misrepresent certain cultures. Additionally, there are concerns around privacy and data security, as AI translation tools often rely on massive databases of user-generated content.

Third, AI translation tools often create phantoms; they “often go off on random tangents, or even go so far as to create fake translations on occasion.” Today’s AI translation tools are, therefore, unreliable, ethically questionable, and unpredictable.

But MMP would have us believe there is something “surreal” about being in “staunch opposition” to AI translation. So, I ask: what is surreal about being in opposition, even in staunch opposition, to unreliable, ethically questionable, and unpredictable tools? And I answer: nothing, nothing at all. On the contrary, opposition to using unreliable and ethically questionable and unpredictable tools to create one’s work product is the most real and understandable and reasonable opposition any service provider can maintain. There is nothing utterly bad or utterly stupid about this opposition. No opponents of AI translation are derelict in the performance of some duty or doing the translation industry some disservice.

Where the proponents of AI translation do not couch their ad hominem attacks in seemingly objective prose, they will just call names. MMP, for instance, calls us “lame-duck translators […] who do nothing but bring our beloved industry into disrepute.” Where these proponents don’t call names, they will typically retort by citing AI translation’s “enormous potential” and its ability to create “potential[] efficiency gains.” And cite as they might, they will typically omit, and MMP in fact omits, two important explanations. MMP omits (1) an explanation as to why anyone should use tools whose potential is enormous, but whose actuality is unreliable, ethically questionable, and unpredictable; and (2) an explanation as to why potential efficiency gains can be expected with unreliable, ethically questionable, and unpredictable tools. I don’t doubt in the least that the future will unlock at least some of this potential. But the future is not the present, potentialities not actualities.

This truism is even borne out by current translation practice. No professional translation service provider uses unreviewed or unedited AI translation output of an entire text. Or rather, if it does, best practice requires that

• such output be used in accordance with instructions by the client, but that the client be advised of the risks posed by that output; or
• such output be used for nothing other than to establish the rough and very general gist of a text, sometimes to establish whether cost or time savings can be had by sending only a certain part or only certain parts of a text for professional translation.

The reason for this very limited use of unreviewed or unedited AI translation output may or may not have to do with ethical considerations, but it certainly does have to do with the very simple fact that AI translation tools are simply unreliable and unpredictable.

Is this limited use of AI translation output the realization of some enormous potential? No. To call it that would be to strain credulity. Is it the realization of a potential efficiency gain? Yes, but with one caveat: it is not the efficiency gain the proponents of AI translation tools would have us believe it is; it is not an efficiency gain in the sense that a true and complete translation – the two metrics of any good translation – is rendered in some more efficient manner on account of some translation tool, AI or other. What is more: by no stretch of the imagination will such very limited use take center stage in all aspects of translation life, as MMP predicts it will.

To date, however, there are two uses of AI translation tools that have taken the stage, but certainly not center stage, in some, certainly not all, aspects of translation life.

The first, and perhaps most ubiquitous, use consists of the piecemeal use of AI translation tools to aid in the translation of certain sentences, phrases, and the like. Typically, this aid comes in the form of an API, which is integrated into a computer-assisted translation tool, a so-called CAT tool, and is activated when a translator is faced with an increased difficulty of some kind: the syntax is strange or unfamiliar, words seems to be used in unfamiliar senses, etc. The idea is that the AI translation output provides the translator with something, anything, and that this something serves not as a suitable solution, but as a springboard to a suitable solution.

And like the very limited use of AI translation output discussed above, this use of AI translation output is the realization neither of some enormous potential nor of some potential efficiency gain in the sense proponents of AI translation would like us to believe. The aid provided resembles a crutch, and a crutch points to an impairment of some kind, a broken leg, a sprained ankle, whatever. It helps people to walk who are having difficulties walking; it’s an enabler, ideally a temporary enabler. In point of fact, it is just as likely to help a translator walk steadily as it is to slow him or her down. It doesn’t offer any translator any enormous potential. It doesn’t help any translator to walk more efficiently. It just helps the translator to walk.[4] That’s the best one can say.

The second use of AI translation that has taken the stage in some aspects of translation life is machine-translation post-editing, so-called MTPE, of which there are two main kinds: light MTPE and full MTPE. Light MTPE is a translation service in name only. With light MTPE, the AI translation output is glossed over by a so-called post-editor, usually someone with a translation background of some kind, to remedy the most obvious of errors – and that’s it. It is used mostly for one-off texts, ephemeral matters of little to no import, gisting, and the like. By contrast, full MTPE is a translation service. With full MTPE, AI translation output of an entire text is provided to a human translator, and that human translator is tasked with turning that output into a true and complete translation – that is, into a high-quality translation.

Of course, the theory is that the AI translation output should expedite the process, should ensure that the translator delivers such a translation faster than he or she otherwise would. And of course the theory stands and falls with the quality of the AI translation output. And of course, the theory turns on empirical evidence. But I, personally, have never seen any AI translation output that would justify full MTPE – and I have seen a lot of AI translation output. What is more, to the best of my knowledge, there exists no reliable empirical evidence that full MTPE expedites the process. At the moment, the best evidence the industry seems to offer is a ludicrously expensive $585, sixty-page report no one will read; a handful of studies with wildly varied results; and the biased word, or personal anecdotes, of a few market players that offer AI translation tools and/or MTPE services – complete with boasts of how fast they can churn out post-edited translations.[5] It might be too harsh, but I’ll say it anyway: the evidence does not read like hard data; it reads like advertising copy.

Given this state of evidence, it is fair to say the jury is still out on whether full MTPE realizes enormous potential or efficiency gains. While I doubt it, I won’t rule it out. And if you or someone you know uses AI translation tools and have realized efficiency gains, perceived, imagined, or real – good for you. Keep on keepin’ on. I do not wish to take that from you.

But let’s keep certain facts straight. AI translation is still in its infancy. It is still too unreliable, ethically questionable, and unpredictable to demand widespread adoption by professional transaltors today. Its general potential, enormous or other, and its potential efficiency gains remain just that; they are objectively potential. Hence, early adoption of this technology remains a personal matter, and its earlier adoption or non-adoption does not make anyone bad or stupid, does not mean someone is derelict in the performance of some duty or is doing the industry some disservice.

I am an opponent of AI translation, for all the reasons given above, but also for the simple reason that I have never seen any good AI translation output. I also feel obligated to keep its faults in the public eye, because AI translation output has improved over the years and these improvements have lulled professionals outside the translation industry – lawyers and others – into a false sense of security. These professionals understand precious little about translation generally, even less about quality assurance in the translation space, and even less still about what makes a good translation good. They should not be denied information about the risks posed by AI translation and its overzealous advocates.

Nor should they be fooled into thinking that good translation work product is determined by the use of the latest tools available. It’s not. It’s not even determined by the use of the best tools available. It’s determined by the best use of available tools. The tools do not a translation make. Personal ability does. Translation is not a product, it is a service.

End notes

[1] Page 201 of Alexander Pope (2006). Peri Bathous, or the Art of Sinking in Poetry. In P. Rogers (Ed.), Alexander Pope: The Major Works (pp. 195-239). Oxford: Oxford University Press.

[2] Page 102 of G. K. Chesterton (2009). Varied Types. In The Wit, Whimsy, and Wisdom of G.K. Chesterton (Volume 6) (pp. 83-186). Landisville: Coachwhip Publications.

[3] AT&T was the iPhone’s exclusive wireless carrier in the United States at the time the iPhone launched in 2007. See Wikipedia’s “History of the iPhone.”

[4] The present author has heard of translators who use this crutch for every sentence of every translation; they are said to use AI translation to see whether it offers anything useful. But this is no longer using AI translation as an aid; no competent translator needs help with every sentence. This is using AI translation for its own sake. And using a tool for its own sake does not actualize some enormous potential. Neither does it represent some realized efficiency gain.

[5] See the Slator article titled “How Fast Can You Post-Edit Machine Translation?” dated December 12, 2022, for a summary of both the studies and market players.

Hinweise zur bestehenden Moderationspraxis
Kommentar schreiben