What does it mean to translate in a world where everything is translated automatically?

We translate constantly, often without even realising it: an automatically subtitled reel, an email drafted with DeepL, a conversation mediated by ChatGPT, a Korean series streamed online. Translation has become invisible, instantaneous and embedded in everyday life.

 Yet long before artificial intelligence and global platforms transformed the way we communicate, Umberto Eco had already recognised that translation is far more than replacing words from one language with those of another. It involves interpretation, choice, loss and, inevitably, a transformation of meaning. It is no coincidence that one of his best-known books on the subject is titled Dire quasi la stessa cosa (Saying Almost the Same Thing). It is precisely that “almost” that captures the full complexity of translation.

 Today, in a world where the internet promises instant communication and digital tools can translate almost any text in a matter of seconds, Eco’s reflections on translation feel more relevant than ever.

 To revisit Eco’s ideas in the age of digital communication, artificial intelligence and global connectivity, we spoke with Barbara Ivančić, Associate Professor of German Language and Translation at the Department of Modern Languages, Literatures and Cultures at the University of Bologna. 

Umberto Eco famously wrote that translating means “saying almost the same thing”. What does that “almost” mean in the age of artificial intelligence?

From the perspective of machine translation, one might argue that the “almost” has become less relevant, because translation is viewed primarily as a process of decoding and replacing symbols. The technological landscape has changed dramatically since Eco wrote those pages, but the basic premise behind machine translation remains much the same.

Eco carried out his own experiments with machine translation before the advent of neural systems, when systems relied on statistical methods. Although these represented a significant improvement over earlier approaches, the results were still fairly poor, as anyone who used the first versions of Google Translate, or even earlier tools such as Babel Fish, Altavista’s translation service in the early 2000s, will remember. It was Babel Fish that Eco used for the experiments described in Dire quasi la stessa cosa (2003). As many readers may recall, the results were often unintentionally hilarious. As Eco pointed out in that book, it was clear at the time that these systems lacked both encyclopaedic knowledge of the world beyond the text and the contextual information needed to disambiguate words and determine their meaning in a specific context.

Today, in the age of neural machine translation and generative AI, the technologies that underpin large language models (LLMs), those same experiments would undoubtedly produce far more convincing results. This is because machines are now designed to learn autonomously, or, in technical terms, through machine learning, drawing on vast quantities of linguistic data used during training. Yet beyond the technical advances, language remains, for these systems, a mathematical construct. More than ever, translation outputs are the result of numerical transformations and probabilistic calculations. The “almost” Eco referred to belongs to a different realm altogether: that of interpretation and negotiation, concepts to which he devoted considerable attention in Dire quasi la stessa cosa and throughout his work. These concepts concern the meaning of linguistic signs, the sociocultural environment in which a text is produced and received, and, not least, the emotional and affective dimensions involved in reading and interpreting a text. All of these elements lie beyond the scope of language machines. For this reason, I do not believe that Eco’s “almost”, at least as he understood it, can be meaningfully applied in this context.

Do machine translation systems actually translate meaning, or merely linguistic correlations?

I’ll answer by quoting the linguist Giuseppe Antonelli, who, in an article published in La Lettura, the cultural supplement of the Italian newspaper Corriere della Sera, on 10 August 2025, reflected on large language models designed to generate outputs across a range of natural language tasks. As Antonelli writes, these outputs are “sequences of words that, when put together, look very much like human language. But they are not.  Rather, they are the product of an entirely alien process of numerical transformations, one that bears little resemblance to the way human beings produce language.” Machine translation, as one of the many tasks performed by these systems alongside text generation, summarisation and analysis, falls squarely within this framework.  The same applies even to models trained exclusively on translation data.

Going back to the previous question, regardless of the extraordinary improvements we are witnessing, the underlying principle of machine translation remains essentially the same as that which guided the earliest research in the field. It is worth remembering that much of this research was heavily driven by military interests during the Second World War and the Cold War. Its primary goal was not communication across cultures, but the deciphering of encrypted enemy messages.

 Alan Turing, now best known for the Turing Test, played a crucial role in this phase, as did the mathematician Warren Weaver, whose name is associated with the famous Weaver Memorandum (1955), one of the earliest texts to consider the possibility of using computers for translation between languages. Significantly, Weaver opens the essay with what he calls a “war anecdote”: the story of a Turkish text given to a mathematician who, despite having no knowledge of Turkish, managed to reconstruct the original message. From this, Weaver concluded that “this process made use of frequencies of letters, letter combinations, intervals between letters and letter combinations, letter patterns, etc., which are to some significant degree independent of the language used”. I believe that observation still provides a compelling answer to the question we started with.

 What is lost when translation becomes fully automated?

From my perspective, quite a lot is lost. First of all, we risk losing our relationship with the very idea of untranslatability. The notion that everything can be translated at the click of a button, an idea understandably promoted by the language industry, distances us from the awareness that not everything can be translated, and that we need to learn how to engage with what remains untranslatable. This does not mean giving up on translation. Quite the contrary. It is precisely because not everything can be translated perfectly that we continue to translate, accepting that something may always be lost in the passage from one language to another, and from one text to another. That is precisely what Eco's “almost” refers to. It reflects a human approach to translation, one that listens to nuance, embraces interpretation and acknowledges ambiguity. Technology, by contrast, is built around speed, efficiency and scale.

 From this perspective, the relationship between translation and technology is both complex and deeply ambivalent. It inevitably forces us to reflect on our relationship with language itself and with linguistic diversity. These issues are explored particularly well in The Routledge Handbook of Translation Technology and Society (2025), edited by Stefan Baumgarten and Michal Tieber, which I would recommend to anyone interested in these questions.

 For me, automation also entails the loss of the embodied dimension of translation. Emotions, perceptions, personal experience and even the unconscious all play a role in how we read and interpret texts. They shape our translational choices, whether consciously or not. That is why, fortunately, no two translations are ever exactly the same. This is not to say that all translations are equally successful. Rather, it means that every translation, like every literary work, bears the imprint of a human presence. These are aspects that will inevitably be transformed by the automation of translation and writing itself. To me, those transformations amount to losses.

 We should also speak of loss in social and cultural terms. The pursuit of immediate translatability raises ethical, political and environmental questions that deserve far greater attention than they currently receive. This is something we ought to discuss much more, especially within academia. We should be asking what society loses when increasing amounts of attention, funding and trust are invested in an industry from which large language models have emerged, an industry built on forms of human and environmental exploitation and on what many have rightly described as a vast appropriation of intellectual property. After all, without the linguistic data acquired by a small number of private companies, we would not be talking about translation automation in these terms today.

What does a society lose when, apart from a handful of dissenting voices, it embraces without question an ideology of instrumental rationality that ultimately serves the interests of the dominant actors in the global economy?

These are the questions that should be at the centre of discussions about machine translation, rather than an almost exclusive focus on the tasks machines still cannot perform. We should also be paying closer attention to the risk that both language itself and our sensitivity to linguistic nuance may gradually become flattened if we feed exclusively on outputs generated from the data on which machines are trained. I like to think that Eco, in his own way, would have helped keep a spotlight firmly trained on issues such as these.

Eco argued that every translation is a form of interpretation. Does the same hold true for AI?

No, not if we understand interpretation as the attempt to grasp a text's deeper meaning and recreate its effect. As Eco pointed out, interpretation in translation goes hand in hand with the negotiation of meaning: the process of deciding what a translation should convey and how it should convey it. This is very different from the mathematical logic that underpins artificial intelligence systems, which operate by performing calculations on numerical representations of data, linguistic and otherwise, as I mentioned earlier.

 One possible objection is that, in machine translation, interpretation and negotiation still have a role to play. They come into play when users provide detailed instructions to a system, through prompting, and especially during post-editing, when machine-generated output is revised and corrected by a human translator. However, this remains a highly debated issue. Post-editing raises a number of questions about how much room is actually left for human intervention understood as interpretation and negotiation.

 In particular, I am referring to what is known as the priming effect, the influence that machine-generated output exerts on the person revising it. A growing body of research shows that, especially in literary translation, this effect shapes not only linguistic choices but also the way a text is interpreted. In both cases, there is a tendency to remain anchored to the machine's initial output. Creating the critical distance needed to evaluate alternative solutions becomes much more difficult. Yet that distance is precisely what allows translators to reflect on their choices and to regard them as the result of interpretation and negotiation, rather than simply corrections applied to a pre-existing text.