In the Pod čarou newsletter every Saturday, Matouš Hrdina describes social trends that we see all around us, but disappear a little between the lines in the influx of daily news. If the sample interests you, subscribe to the full version of the newsletter.
Technological development in recent years has repeatedly confirmed the rule that even the most revolutionary device or service can fail completely, if a truly practical use for a larger number of people is not found for it, which is not just an inflated speculative bubble.
It was evident with blockchain technology, even Mark Zuckerberg’s dreams of a metaverse vanished like steam over a pot, and we could endure examples for a long time – recently it turned out that Apple’s Vision Pro smart glasses were not such a blockbuster again, completely as expected by critics, and the company had to reduce plans for their production.
The debate about real usability also revolves around new AI tools, but it is somewhat more complicated. There are many different very useful options for deploying AI, but investors have much higher expectations and thus focus mainly on utopian visions that will probably never be of real interest – this is well described for example https://twitter.com/doctorow/status/1782675067861614933?t=VvYY8_eDn7krku-lhMPY2w&s=19.
Even in this newsletter, I mostly focus on criticism rather than praise of AI tools. But when I recently wondered if there was any significant example where the benefits of AI deployment significantly outweighed the negative social, economic and environmental impacts, I realized that we had already encountered it long before the rise of services like ChatGPT or Midjourney. Many of us use machine translators almost unnoticed every day, and we don’t even realize what a revolutionary technology this is.
Twilight of the translators
Various automatic translators have been around for years, but their practical utility has long been dubious – you may remember how imperfect translations used to be from Google Translate, which was first introduced in 2006. But then machine learning methods began to significantly improve in this area, translators were improving, and when the DeepL tool appeared in 2017 (with translations into Czech in 2021), it was clear that translators were becoming a truly revolutionary technology.
Listen to the audio version of the newsletter read by the author.
The capabilities of DeepL were then analyzed by my colleague Pavel Kasík, and most of the predictions mentioned in his article have been fulfilled since then – it seems quite self-evident to us that we can immediately translate even more complex texts in very decent quality using DeepL or other tools whose quality also got to a very high level.
The possibilities have expanded even more after the improvement of the tools for converting sound to text and vice versa, and we are slowly approaching the utopian concept of a universal translator that we previously knew only from Star Trek – elegant devices like the recently widely discussed Humane AI Pin do not work very well yet, but if we are satisfied with the more cumbersome interpretation using the smartphone screen, language barriers not only on holidays in exotic locations are falling in a way never seen before.
Every new technology also has its dark side, and as is the case with the introduction of AI tools, the first and most direct threat to translators is the disintegration and decline of entire professions due to automation. Statistics from the British Union of Writers, Illustrators and Translators show that more than a third of translators have already lost their jobs due to the advent of AI, and their services are also falling in price. Similar information is also provided by data from the Upwork platform, which offers freelancer services – over the past two years, the advent of AI, in addition to writing texts and customer support, has had the greatest impact on the translation sector, where the number of orders and the amount of fees paid have significantly decreased.
Similar effects can be seen from the above data, for example, among illustrators. However, even though in this newsletter I mostly stand up for the rights of workers in the face of ruthless automation, in the case of machine translations, it is necessary to mention a significant difference.
A large part of the new AI tools offers and imposes options and services that have not yet been in great demand among ordinary individual users – typically, for example, the creation of AI illustrations, which we were without until the advent of Midjourney et al. they worked without any major problems. In this way, they primarily benefit companies and entrepreneurs, who thanks to them no longer have to pay experts and arrange the required services themselves, albeit of a lower quality (and in addition through AI tools, which mostly used the works of human artists illegally and for free as training data, and thus exploit them twice).
However, translations are a rarer example of a service for which universal interest among ordinary users undoubtedly existed, because everyone simply needs to translate something from time to time. Therefore, they have become an extremely useful and widely used tool and a regular user who, with two clicks, has a website translated from Spanish to Czech, in short, we cannot so easily accuse them of exploiting other workers (unlike a company that hires a human illustrator, filmmaker or copywriter replaced by a machine).
The speed and naturalness with which we have all become accustomed to the possibility of automatically translating anything immediately shows that the twilight of translators – at least in the sense of ordinary translation of banal utility texts, not fiction or poetry (more on this below) – is really inevitable in this case . Not because of the greedy corporate clientele profiting from automation, but because the new technology really brought a fundamental benefit for everyone and almost no one wants to go back to the times when the manual for the washing machine had to be translated by humans.
The illusion of global intelligibility
However, the advent of machine translation also brings with it other, less visible risks and negative impacts, which will probably only get worse. First of all, there is a danger that people will fall into the illusion that there is no need to learn foreign languages anymore and will start relying only on the help of a smart translator algorithm.
The data shows that this is not an idle concern. In an interesting analysis for The Atlantic, Louise Matsakis recently pointed out that the number of courses taught in languages other than English has declined massively in American universities over the past ten years. To a large extent, the underfunding of the humanities and the typical lack of interest of Americans in foreign languages are probably to blame, but the influence of machine translation cannot be ignored – data from South Korea, for example, where interest in studying any other foreign languages beyond English indicates this.
Another, no less serious, risk is the fact that machine translation continues to reinforce global linguistic inequalities, endangering smaller languages and confirming the dominant role of English. Large language models work best in English, and most translators or text generators focus only on the ten largest world languages, where there is the largest amount of quality material for training datasets and also the most potential users.
More than half of all websites are in English, which is not spoken by 80% of the global population, and the ability to easily generate new websites and texts is likely to exacerbate this disparity. Small languages, especially from poor developing countries, which already suffer from the consequences of colonialism and global inequalities, are at a disadvantage due to the lack of high-quality digitized texts that can be used for building language models. And a fatal disadvantage is the fact that giving instructions to AI tools (so-called prompting) is most effective in English, which in turn further widens the gap between dominant and minority languages.
Perhaps the most insidious threat is the transformation of our behavior and expectations with which we approach interpersonal communication. In the aforementioned analysis, Matsakis points out that thanks to machine translators, we have come to automatically expect that a message in any language will be readable for us and that we are entitled to articles or films in our mother tongue. And from that it is only a step to the foolish belief that everything is therefore automatically fully understandable to us and nothing is “lost in translation”.
This is of course absurd. Language is primarily for interpersonal communication, not for communication with a machine, which is exactly what we do when working with an automatic translator. There are many hidden meanings, metaphors, imperfect synonyms, customs and cultural factors that come into play in a debate with other people, which the algorithm cannot capture. Everything gets even worse if we move from the written text to the spoken word, where the tone of voice, the relationship with the speaker and a whole host of other influences come into play (which is why interpreters don’t have to worry about their work nearly as much as translators).
This can be seen every day, for example, on TikTok, where a number of creators who are not native English speakers help their videos with automatic English dubbing, but the sentence structures of which reveal their roots in a completely different language, and everything then seems rather convulsive. Just because someone gets the gist of your message through a translation doesn’t mean they actually understand it, let alone interpret it the way you intended – another example of this illusion was the viral campaigns after the beginning of the Russian invasion of Ukraine, when Western users started with the help of automatic translators into Russian, flood the Russian Internet with various polemics and information about the real state of affairs, understandably with zero impact (and it can be added that Russian trolling on social networks in the opposite direction is also largely enabled by machine translators).
Holes in the perception of the world
Even if machine translation were to improve itself, it would probably never be able to deal with one of the fundamental principles of our language use. The theory of linguistic relativity (sometimes also known as the Sapir–Whorf hypothesis) has resonated in linguistics and other social sciences for more than a hundred years, which claims that language fundamentally affects our perception and understanding of the world around us.
Scientists argue about how and to what extent this phenomenon really works, but even on a layman’s level it is obvious that we simply function differently in different languages and are a slightly different person. Some dimensions of existence are richer in some languages than in others (this nice gloss on the use of dual number in Slovenian, which cannot be directly translated into most other languages, shows this), and if we want to convey them through translation, it requires the careful work of a professional human translator . In short, in addition to knowing the cultural context and unspoken meanings, the latter thinks about language in a completely different way than a machine operating only with statistical calculations of optimal translation.
Words have a different meaning than the one we study in the dictionary, and we should not be under the illusion that machine translators are a universal solution to all problems with language barriers – just read the fascinating write-up of the work of the translator Jennifer Croft, who translates into English at length the only paragraph of Olga Tokarczuk’s extensive novel The Book of Jacob.
If we resign ourselves to the study of foreign languages because of the tempting availability and simplicity of machine translation, we will completely lose the opportunity to look into other and fascinating dimensions of reality and its perception. But we will lose them even if machine translations between several large languages start to dominate the Internet, while the smaller ones fall into oblivion. And last but not least, we will also lose them if we turn professional human translation into an outclassed and seemingly outmoded profession.
In this area, it is a sad truth that progress is inevitable, the use of machine translators (in contrast to the vast majority of other AI tools) can be characterized with reservations as a contribution to the common good, manuals for irons will only be translated by machines and translators will irretrievably decrease . Intoxicated by new possibilities, we should not forget the fact that the “thinking” of the algorithm has its insurmountable limits, and even if it can facilitate basic interpersonal communication, we still have to arrange real understanding ourselves.
If you liked the example from the Under the Line newsletter, subscribe. Every Saturday you will receive it directly in your e-mail, including tips for further interesting reading from Czech and foreign media.
