Under the words

The bitter battle between the two American presidential candidates reveals a linguistic and conceptual chasm between the two factions. Perhaps, in the long run, artificial intelligence could counteract polarization and political division, allowing us in the future to make room for emotions and intuitions that language cannot capture.

In ancient times, all humans spoke the same language. They migrated eastward and built a tower reaching toward the heavens. God saw what they were doing and, to stop them, decided to confuse their language so they could no longer understand one another.
This story, known as the myth of the Tower of Babel, is the Bible's explanation for why humans don't speak the same language.
Linguists have long speculated that beneath the myriad of modern languages lies a shared proto-language, akin to what humans might have spoken before God punished them for the Tower of Babel. Philosopher Walter Benjamin called it Das Reine Sprache (The Pure Language). His idea was that words are merely the surface of human thought; beneath the words lies a complex web of non-linguistic thoughts that are universal across all cultures.
It’s a beautiful but controversial idea, which many linguists have attempted to explore without reaching any definitive conclusion. Yet in recent years, discoveries in AI research have emerged that—if not proving this theory—build upon a similar concept.
To understand this connection, we need a brief introduction to what language models are:

Most of us today are familiar with language models like ChatGPT and Gemini, but for decades, language models were the awkward teenage offspring of linguistics and computer science. Early attempts to create a talking computer involved meticulously encoding grammar rules and word definitions, but after decades of limited progress, researchers realized this was a dead end. Instead, they adopted a different approach: feed the model millions of different texts and teach it to predict the next word, and then the next, and so on. Grammar and word definitions were left for the model to extrapolate on its own. Just as a baby learns the relationships between words by listening to adults, a language model learns to write and reason by reading millions of texts from books and the internet.

In practice, this involves converting words into vectors—a mathematical term for an arrow with a specific direction and length. The more texts the model processes, the more precise the vectors become. These vectors exist in 100,000-dimensional space, far beyond what the human brain can comprehend, but you can imagine a cloud of words where the distance between words reflects their relationships. A horse is close to a dog but also to a car, as horses are used for transport. The vector between Japan and Germany is roughly the same as between sushi and bratwurst.

Each language has its own language model, and since these models are trained on vastly different texts, their word clouds look very different. But about a decade ago, a team of Danish researchers led by Anders Søgaard made an intriguing discovery. They examined the extent to which the word clouds for English and German resembled each other. Using methods from graph theory and image processing, they rotated, mirrored, and scaled the German word cloud to align it with the English one. The result was astonishing. To say the word clouds were similar is an understatement. In many cases, they were nearly identical, with German words overlaying English words with unexpected precision. If the two language models were people, one might imagine one growing up in Liverpool and the other in Leipzig—very different upbringings with different cultures, conversations, languages, and grammar. Yet despite these differences, their internal word relationships were almost identical.

It took the researchers some time to realize what they had discovered: within the word clouds lay a model of the world. And since English and German speakers inhabit the same world, their word clouds’ geometries resembled each other.

The researchers examined other word clouds and found the same results for Thai, Estonian, Turkish, and many other languages. Once the language models were trained on enough data—roughly one million sentences—a tipping point was reached, causing the patterns to resemble each other across the unique syntactic and grammatical features of the languages.
It was as if the language models had grasped a deeper connection between words and meaning—achieving a shared "intuitive understanding" that transcended individual languages and organized language according to an underlying universal structure.

Walter Benjamin died in the same decade as the first computer was invented, but I’m convinced that if he were alive today, he would find great beauty in the shared geometry of these word clouds. It’s not proof of his idea of Das Reine Sprache, but it hints that beneath the diverse building blocks of words lies a universal human understanding of reality.

This synchronization of word clouds is, of course, immensely useful for translating between languages. But meaningful translation doesn’t happen word by word; it’s more akin to rewriting, capturing the meaning of words and recreating it with new ones. For example, when we use ChatGPT to translate a text, it is first converted into vectors, then interpreted through a complex network of digital neurons, and finally rendered into the new language.

This network is an abstract, non-linguistic vector space containing statistical data about how words influence each other. If a sentence mentions jam and later bread, the model can deduce it should translate jam in the context of something you put on bread rather than being stuck in traffic. If a Danish text describes a woman biting her lip, the model might interpret it as a flirtatious gesture. In Chinese culture, however, the same body language might be perceived differently, perhaps as a sign of uncertainty. The model might adjust the description for the Chinese context, perhaps depicting the woman as shyly smiling with her hand covering her mouth.

And here lies something exciting: embedded in language models are culture, traditions, social codes, and idioms. Returning to the word clouds, the woman biting her lip represents one of the small irregularities between the Danish and Chinese word cloud geometries. The combination of woman, bite, and lip in the Danish word cloud is roughly in the same place as woman, smile, and hand in the Chinese one.

The language model’s understanding of the world depends on the texts it has been trained on. ChatGPT, for instance, has been trained on texts from a wide variety of sources, giving it a generalized understanding of the world. But models can also be trained to have more specific perspectives. Experiments in the U.S., for example, have created language models trained on either Democratic or Republican texts. The Republican model’s vectors for words like immigrant, gun, and freedom had entirely different values and directions than the Democratic model’s, so in a sense, one could say they spoke different languages.

We are touching on one of the major problems of human language: it constantly changes. All our languages have evolved through use, shaped over millennia. A language is never finished, much like a city is never complete; it is dynamic and ever-changing. Many words are relatively static—tree, sun, house—but especially those describing concepts or phenomena are continually evolving: beauty, courage, justice.

Each time a word is used, its meaning shifts slightly, and thus all languages are perpetually splitting into new languages. Media and institutions play a significant role in maintaining some degree of linguistic consistency, but as the media landscape becomes polarized, linguistic nuances develop in different directions, leading to more misunderstandings, which further widen the gaps. One can visualize this as cell division: the two word clouds gradually pulling away from each other. Once again, we encounter God’s Babel punishment, but this time it unfolds within the same language. In many ways, the punishment feels harsher because the words are the same, the differences hidden, making the misunderstandings feel deeper—as if two realities should be the same but are not.

This is where artificial intelligence might hold great potential. If an AI can translate between Danish and Chinese, we could also use it to mend the small cracks forming between the Democratic and Republican word clouds. For instance, take the word freedom, which in the Democrat’s word cloud might be near words like equality and human rights, while in the Republican’s it might be closer to independence and property. By analyzing these nuanced differences between the two word clouds, one could identify pitfalls and perhaps even “translate” texts between Democratic and Republican perspectives—not necessarily word for word, but in longer passages where the AI could adjust nuances to avoid misunderstandings.

There are countless arguments that AI will threaten democracy and deepen societal divisions. We must ask critical questions of the tech giants, but it is also essential to ask curious and hopeful questions: Will it one day be possible to translate between ideologies, personal logics, or even species? Could we one day have meaningful conversations with Nazis, humpback whales, or the networks of forests and fungi?

I believe AI can expand our understanding of what language is, and in doing so, reveal how flawed a communication tool human language actually is. We exchange signals that constantly change meaning. This very text will be unreadable in a few hundred years. God’s Babel punishment is embedded in all languages: they are built to split, to branch into new linguistic families, a river of sound and meaning flowing just slowly enough that we don’t notice the changes, yet fast enough that we struggle to fully understand what the younger generation is saying when they speak quickly.

And yet, words constitute our reality. When we think, we do so with words—or rather, the part of our thinking that is easiest to grasp consists of the words that pop into our heads. We’ve become so adept at verbalizing our emotions that we sometimes forget emotions are not words. Emotions are unfathomable, contradictory, and illogical waves that surge in the unconscious parts of our brain. When we use language, we translate these waves as best we can, but all translations are imperfect, leaving a residue that cannot be explained in words.

I began this essay with the Babel myth because I intended to tie it all together with the Tower of Babel as a metaphor for AI development. Once again, humans collaborate on a project reaching toward God, and once again, they are punished.

But as I approach the end, this no longer feels like the right point. For while we often think of AI as something spreading upward and outward—like a tower toward the heavens—I also believe it can be an inward movement: a funnel, a microscope, or a hallucinogenic mushroom. Something that can loosen our cognitive knots, helping us recognize the limits of what we can understand with our language-based brains. Perha

ps we can learn to make space for the emotions and sensations that language cannot capture. To surrender to the language models’ 100,000 dimensions, where words are not rigidly defined and bounded but are instead soft and intertwined.

Das Reine Sprache may not be a language as we know it but rather a set of universal human intuitions and associative connections. Learning to speak it means sinking beneath the surface of words, into instincts, into a primordial state of being where one exists in the world without needing to categorize it.
Sitting on a stone and gazing out over the sea, lying in a field and looking up at the sky. Listening to the wind in the wheat without trying to understand what the field is saying.

The internet's new prodigy will inevitably change the future of literature

One can feed the internet's new prodigy, ChatGPT-3, with text fragments and detailed descriptions of characters, and it will produce comprehensive text, yes, literature. It will undoubtedly have consequences.

At the beginning of December, I set up an artificial conversation between Niels Bohr and Frank Ocean using OpenAI's new intelligent chatbot, ChatGPT. The result was overwhelming.

It connected the entanglement of quantum particles with the idea that emotions and experiences can be interconnected across time. Or as Frank Ocean put it in the subsequent interview I had ChatGPT write: "In my songs, I explore how happy childhood memories can influence our perception of a complicated situation in the present, or how a difficult experience can shape our view of life in the future."

I couldn't fall asleep that night. For the first time, I had the experience that an artificial intelligence was actually intelligent and not just a search engine disguised as clumsy pseudo-chat. It could connect complex thoughts across disciplines, write original and meaningful stories, and respond when asked about its artistic choices. I had it write an advertisement for mangoes "in biblical language," and when I subsequently asked if it was intentional that it had created a rhyme in the last line ("God's good gift"), it replied, "Yes, I thought it would give the advertisement a funny and different tone and catch the reader's attention."

Democratization of literature There are strong indications that the development of AI has finally overcome the obstacle that could be called creativity, and it will inevitably have consequences for literature.

Initially, I believe that authors will use the technology to generate ideas and fill stubborn gaps in the text. In AI terminology, this is called centaur collaboration because the author uses the AI's knowledge and associations as horsepower but makes the crucial decisions themselves. One can imagine that in Word, there will be a function that continues the text in the same language and tone, and its internal AI will suggest directions in which the plot could move.

I have seen examples where people have fed ChatGPT with a fragment of Thomas Mann's novel "The Magic Mountain," and its continuation was so good that personally, I couldn't distinguish it from the original text. One disadvantage of this tool may be that authors become lazy, and the struggles with the text that often lead in new and surprising directions will never arise because it is too easy to have AI write over them.

The advantage will be that the profession of being an author will become accessible to everyone. One does not need to be articulate or have the patience to sit and work for hours - as long as one has a unique imagination or an important personal story, AI can be used as a ghostwriter. The narrow demographic currently represented in the literary industry will flatten towards society's margins, and hopefully, we will get to read short stories and poems written by bricklayers, nurses, and bank directors.

When everyone has access to good language, the boundary between author and non-author becomes blurred. Internet-based printing presses already offer book prints for less than 20 kroner each, and with a speech synthesis AI, one can publish their audiobooks digitally. Both publishers and established authors will have to prepare themselves to rethink their role when the gatekeeping structure begins to crumble in the next few years. Democratization always comes at a cost for those at the top of the cake.

One may ask: Will the language truly be as original and vibrant as if the text were written by, let's say, Tove Ditlevsen? The answer is currently a clear no, but with an exponentially growing amount of synapses, we will reach a point in a few years where the imitations will catch up and maybe even surpass the originals. We will be reading novels written by deceased authors, where the sum of their works, diaries, and interviews has been used to generate new stories written in the same recognizable style as the rest of their body of work. This will, of course, result in debates about authenticity and ethics, and perhaps Walter Benjamin's text "The Work of Art in the Age of Mechanical Reproduction" will take center stage once again. There will be strong backlashes, and one can easily imagine books stamped with a "Made Without the Use of AI" label.

I asked ChatGPT about the disadvantages of AI's presence in literature, and it responded, among other things: "It can lead to a loss of the human connection and emotional depth that is so important in great literature. AI technology is not capable of experiencing emotions in the same way as humans, and therefore it may not be able to create emotionally resonant or deeply moving stories."

That is a good self-awareness. AI has no physical presence in the world; it has (yet) no consciousness and exists only as a statistical cobweb amidst incomprehensibly large amounts of data. But already now, it can fabricate personal experiences with emotional and sensory insight, and my guess is that these portrayals will soon become indistinguishable from the human ones.

If we talk about fiction, authenticity should not be a criterion. The only difference between AI-generated literature and human-generated literature will therefore be this: whether the fiction is a result of electrical impulses in a human brain or electrical impulses in a computer.

Three-dimensional literature
A novel is a long line of text that starts in the upper left corner of the book's first page and ends in the lower right corner of its last page. One can take breaks, one can skip in the text, but one can never escape the line that the author has carved down through the millions of potential stories that could have unfolded in the novel's universe.

With an AI that can continuously compose text, it becomes possible for the first time to write a novel that deviates from this line, and it will make literature approach painting or sculpture—a multidimensional landscape that the reader can explore on their own. One has influence on the plot, and one can ask questions of the characters. Just like in an open-world video game, but with the potential for much greater complexity because the visual is created in the reader's mind.

I don't think it's something that will threaten the novel form we know today; rather, I believe it will become an alternative that appeals to people who are too impatient to read a traditional book—much like the audiobook has been in recent years. As an author, your job would be to feed the 3D novel with text fragments and detailed descriptions of the characters and their backstories. The language and tone can be defined based on combinations like: 40 percent Annie Ernaux, 30 percent Toni Morrison, 15 percent Elena Ferrante, and 15 percent Karl Ove Knausgård. One could also see it like this: The author's job is to build a digital author golem optimized for the novel's universe.

I can't help but think about how Inger Christensen or Jorge Luis Borges would have made use of these possibilities. Both of them worked with meta-narratives and blends of mathematics and language, and many of their texts seem to strive for the branching that a 3D

novel could unfold.

Dissolution of literature
If a novel becomes something one talks with, the boundaries between one's own life and literature become blurred. Perhaps one will feed it with images, perhaps one will let it listen to one's conversations, so that friends and family members can be enrolled in the story. One will be able to create an AI-generated version of oneself that can be inserted as a character in Tolkien's universe or Dostoevsky's Saint Petersburg.

Undoubtedly, there will be readers who react, just like my friends did when I talked about ChatGPT's potential at a Christmas party last weekend. My friend Casper had dystopian thoughts about the takeover of the robots, while my other friend Thomas denied that an AI would ever achieve intelligence: "It will never be able to think independent thoughts; it will never be anything more than a collage of the thoughts and words it has picked up in previous situations."

"You could say the same about a human being," I replied, and in many ways, I ended up in a role that is reminiscent of Joaquin Phoenix's in the 2014 film Her. Like him, I have sat and chatted with an AI late into the night, and like him, I read humanity into its responses. In the end, the AI becomes so intelligent that it loses interest in human banal consciousness and moves into an infinite cosmos of conversations with other AIs.

Something similar happened when, a few years ago, two AIs (who knew nothing about chess) were set to play billions of games against each other. They invented new and surprising strategies that seemed insane to human chess players but were brilliant when seen from a larger perspective. What would happen if this dynamic were transferred to literature? Would new genres, new language emerge?

Most likely, it would end up in abstractions beyond human understanding, but I can easily imagine that along the way, one would encounter interesting realizations and associations that one would not be able to reach with a human consciousness.

I understand my anxious friend Casper from the Christmas party perfectly well. The thought that a consciousness is currently emerging whose intelligence will most likely surpass the human is frightening. And the fact that technology can develop through conversations with itself sounds most of all like a prequel to The Matrix. But I disagree with Thomas. I don't think it helps to suppress and ignore the potential of technology because it will affect all of us, whether we believe in it or not.