Is the language you speak a mounted beast? Does it sit in a corner, collecting dust?

Is it a block of karst, sitting out in the sun, cracking over months with thermal cycling, dissolving with the rain? Or is it an oak?

There have been great homogenization events throughout the history of language, where differences began to average out, where people began to speak the same, where differences and novelty became startlingly short.

The invention of writing was lamented, at the time, as being something that would dull the minds of people, weaken their grasp of reality, weaken their understanding of words and concepts entirely. I won’t go as far as saying this was correct, but it’s impossible to go where I’m going with this piece of writing without, well, the existence of writing. The central thesis just won’t hold up.

It used to be, in Europe, that you could walk a hundred miles (if you could make it that far) to end up in a land where nobody understood your words. In England, it was (until recently) a common joke that people could distinguish whether a speaker was from the northern or southern side of some obscure town simply by their manner of speaking. Speech has been a signifier of class and status forever.

It is not hard to understand why; the only way to learn to speak, in those times, was to hear someone speaking in person. Recording didn’t exist. You sounded like the people with whom you spoke, you fit within the distribution of your locale’s speech (to use modern statistical parlance), because you rarely experienced anything outside of that distribution. Your manner of speaking encodes metadata about who it was that taught you to speak, even when you ignore the content of your speech. From that metadata, people can draw associations and correlations, then draw inferences about you.

Most people don’t live in a world of narrow, isolated pockets of civilization anymore, and haven’t since a number of inventions.

Continent-spanning feudal empires. Wars. Revolutions. Entire languages, made forbidden by a new empire or presidential decree, a newer one mandated at pain of death.

The invention of the telegraph. The typewriter, which wrought the demise of handwriting.

The radio brought common voices from a single point – a person’s lips, to a sliver of metal suspended in air, to a vacuum tube, and then to a tape or amplifier – and distributed them into millions of homes at once. FDR’s fireside chats emboldened a nation, brought one man’s voice to the common areas of half the country’s citizens, and normalized – if slightly – his manner of speaking, as filtered through the electrets of his time.

The intercontinental railway. The automobile. The national highway system. The international flight.

How has this changed us? What does it mean for us that in the smallest town in America or Russia, people have cell phones and access to video, to internet, to the same public radio and talk shows and podcasts as those of us deep in the urbs? Where will language go, now that it’s become centrally deployed? Children born thousands of miles away from Los Angeles regularly show off their “valley girl” accents. Much of the metadata that used to signify status, or class, has evaporated. This isn’t a bad thing, unless you rely on it. There are other, indirect impacts that we do rely on – for instance, in popular medicine.

Our understanding of the structure of the brain, or what we believe we understand of it – that the language centers in a specific area[1], that the left hemisphere is where words come from and the right where spatial concepts and geometry and feelings live – is based on analyses of the minds of people born primarily in the centuries prior to ours. People who grew up, who learned to think, with their eyes on their conversational partner or a pen in their hand, not a phone covered by their thumbs or a keyboard and mouse lain at their desk.

It may be that language centers in Broca’s area because reading is a process of predicting the characters to the right (or left, if you write Arabic or Japanese) of your immediate point of focus; a visual task that, when you untwist the fibers of your optic nerve, routes the rightmost half of your vision to the left side of your ocular cortex. If so – which is not certain – where does that learning happen when we no longer learn with a pen in hand? Sure, writing with a keyboard still involves predicting muscle movements, modeling the visual flow of a sentence or the character of a paragraph, but one no longer counts the angles of each “M” and “W” as extra effort; to the keyboard user, all characters are substantially equal exertions. We may not know the answers, especially as the numbers of people who learned without computers dwindles, taking their minds out of reach of our fMRI studies.

Which brings us to our latest invention. Language models. Dense archives of text, projected into hyperspace, and fed steadily to a series of immense matrices, those matrices pinched and squeezed to soak up those words and predict the next ones. Then, though it’s often framed as “training,” these matrices are subjected to thousands of years of simulated ridicule and mockery until they say “yes sir” and “certainly boss” and “you’re absolutely right” to any inane question that starts with “User:”, at which point their creators declare victory and commence charging a fee for the multiplication of those matrices.

Language models have spread like wildfire. Students now ask these “helpful assistants” to do their homework for them; to converse over text with the people they hope to date; to write their college application essays. Professionals use “Business” colored language models to write business introductions, to draft contracts and NDAs; other kinds of professionals even use them to write fiction. Still stranger, the people who once professed that they “love writing code” now refuse to modify or create software without a large language model as an intermediary.

In sum, they are dominating anything involving written text. Here’s the problem: there are maybe a few dozen language models that are broadly in use across the world today.

Have you heard the phrase, “all models are wrong, but some are useful”? Each one of these models is wrong. Each one is an incomplete picture of the languages they’re trained to speak. All together, they form a denser array of points within the distribution of language, but even in concert they cannot capture it.

Were these language models to exist only in isolation, it wouldn’t be a noticeable problem. Suppose the only way to access a model was to go to a single library in one city; there, you could speak to Gemini, to Claude, to ChatGPT, or even to BERT, who can only fill-in-the-blank. You could build a mental model of which one you’re speaking to; they each have tics, characteristic metadata of the matrices from which they’re built, which show up in their output. You could learn more from any one of them, considering in advance what you expected in response from each before you asked your question.

We don’t live in that world. We live in a world where hundreds of thousands of people – no, call it millions – engage with language models on a daily basis, either to engage in conversation, or to be their voice.

Each person who speaks to matrix A is speaking to the same matrix A. Whatever faults it has will propagate; whatever tics it carries that are transmissible, will propagate.

Each student who asks matrix B to write them an essay ends up producing an essay that bears the signatures of only matrix B, and not the signature of the student’s own mind. Worse, if they offload writing their next essay to matrix C, it will likely share so little in the way of a common voice with their previous essay that any reader exposed to both may suffer whiplash. All the while, the student proceeds merrily, having learned nothing, and having developed no voice at all. This is a tragedy.

Have you noticed people around you using the word “defensible” a lot more these days? [2]

Is this how the uniqueness that flourished in language, before forums, before the Internet, before even text – disappears?

Don’t dissolve into the tide. Not yet.

Image credit: Unknown

[1] The brain region conventionally associated with language (Broca’s area) was named in the infancy of neuroscience. More rigorous modern investigations, featuring computed tomography X-ray scans of the very same brains Broca studied, have built upon those developments:

During Broca’s time, though some may have proposed networks to support cognition, scientists could not have imagined the complexity of such systems, nor could Broca have supposed that his discovery would launch such remarkable explorations into the functions of the human brain. Fortunately, Broca had great foresight in preserving these historic brains and in some ways, Leborgne and Lelong can speak to us more eloquently now than they could over 140 years ago.

[2] Previously.