In the Book Genesis in the Old Testament, God saddles humans with a myriad of languages, so they no longer can understand each other, cooperate, and thus avoid God’s punishments. In the New Testament, Pentecost is associated with the miracle when everyone understands each other. And in the cult sci-fi series Star Trek, the Enterprise has a computer system that can translate languages on the fly.

The human fascination with language barriers and how they can be overcome is by no means limited to religious texts or popular science fiction. It has been a focus of computer science almost since the beginning of digital computing. In the 1950s, the US government spent hundreds of millions in today’s money on machine translation, in the hope that this could shed light and meaning on documents obtained by spies in the Soviet Union. The effort failed.

Two decades ago, Google Translate launched as a free service to help humans translate text from one language to another. I remember these early days, and it felt almost magical. It was far better than the stilted, bloodless texts produced by existing machine translation tools. Surprisingly perhaps, Google Translate did not actually understand any language; it had no sense of meaning. It was based on probabilities of word sequences, trained by analyzing billions of lines of text. At one point, Google even established a kind of statistical meta-language, enabling the translation from one language into any other, even if little direct translations between the two languages to use for training existed. The hope was that we could have a single statistical framework, so that each sentence in one language could be translated easily into any other.

It’s the powerful idea of Pentecost: we all understand each other. Because there is only one perfect translation — and the machine will deliver it. Machine translation will produce comprehensive understanding and result in eternal peace. As we know, that did not happen. It may be an enticing idea for some, but it is also a preposterous one. Humanity evolved language as a powerful means to coordinate and cooperate. Language enabled social learning, so that we could, figuratively speaking, stand on the shoulders of others, benefitting from their insights and expertise. But as language evolved, its richness grew as well, producing multiple words and phrases with somewhat similar meaning for humans to express themselves. As a result, there is no single language of truth (apart from religious miracles for the true believers).

Truth is fleeting the moment we express it through words. Because recipients must fill these words with sense, give meaning to mere symbols. And as they do, they open Pandora’s box of interpretation.

The great Umberto Eco, a learned semioticist, believed in the multiplicity of meaning, the multi-layered reality of language. To him, textual interpretation opened a vast open plane, but it wasn’t infinite. There were limits to how far meaning could be stretched when reading text. Others disagreed. Literary critic Stanley Fish saw the interpretative universe as practically unending. But everyone agreed that most texts have many different meanings. If that’s true with text in one’s own language, it applies doubly to the act of translating text from one language, and its social and cultural context, into another. In fact, some literary texts come with so much cultural connotation that the footnotes describing them seem longer and more numerous than the text itself. Just look at a translation of James Joyce’s Finegan’s Wake!

Our most recent digital tools embrace this textual ambiguity. Large language models, like ChatGPT, are perfectly capable of offering multiple versions of translated text. Given some training text, they can mimic style and flair of the original author even in another language and do it astonishingly well. Just ask an LLM to translate text and write it in the style of Ernest Hemingway or Emily Brontë. Here, too, the machine does not have a sense of meaning, it only mimics the probabilities of certain words and phrases used by specific authors. The result is so impressive for us humans because we take these words and put them in context.

Having a tool at one’s fingertips that can translate in different styles and mimic vastly different ways of expressing the same, is not just impressive. It drives home the point that translation is no exact science, but a deeply human art. And that meaning is not contained in words like vessels hold water, but that concrete meaning emerges only as a recipient reads (or listens) to the words. “Lector in fabula” called Umberto Eco his book on this process. As machines now can produce manifold versions of translated text, as we can tweak tone and style to influence meaning with ease, and as we can produce subtle variations of meaning, have we finally conquered the challenge of language barriers and enabled comprehensive understanding?

Translation is no exact science, but a deeply human art.

I suggest not. In fact, I contend that what truly has been pushing us forward as humanity isn’t primarily the understanding of what others said in a different language, but almost the opposite: the misunderstanding of text. It is this linguistic grist that drives discussion and leads to insight. It is the little sand in the well-greased machine of translation that prompts us to question, to converse, and to deliberate. Misunderstandings ignite and kindle the fire of debate. They help us investigate beyond the core of a concept and push it to its limits. Misunderstandings are the catalyst to explore a point beyond its author’s immediate intention. For open minds, misunderstandings are windows through which the light of interest in the diverse sensibilities of humans enters, and through which people can see what lies beyond literal expressions.

A millennium or so ago, the space between words was introduced in written text and the period at the end of a sentence. It enabled a monumental shift. No longer did humans have to read text together and aloud because it was (without spaces and periods) so difficult to decipher that one needed a co-reader experienced with this very text as guide. With space between the words, humans could read silently, and thus at their own pace — reflecting on what they read, rereading if necessary. It freed reading from the shackles of conformity. Misunderstanding does the same for translations. It enables us to think beyond, as we question and argue. Overconfidence in machine translation could hide some intercultural misunderstandings: questioning and arguing proves to be a healthy practice.

I contend that what truly has been pushing us forward as humanity isn’t primarily the understanding of what others said in a different language, but almost the opposite: the misunderstanding of text.

But — and here is the rub — misunderstanding is something deeply human. Because it isn’t statistical, it is rooted in actual interpretation, contextualization, and the search for meaning. And thus, is elusive to technical tools that operate on probabilities.

Perhaps then, there is an obvious symbiosis emerging between humans and machines when it comes to translations: that machines translate while humans (mis)understand. Let’s hope so, because it would be catastrophic for humanity if it were the other way around.

Viktor Mayer-Schönberger

Professor at Oxford University

Viktor Mayer-Schönberger is professor of internet governance and regulation at the University of Oxford. His books include Framers: Human Advantage in an Age of Technology and Turmoil, with Kenneth Cukier and Francis de Véricourt; Guardrails: Guiding Human Decisions.