Table of Contents
The Strange New Grammar of a Half-Machine Tongue
Something quiet is happening to the English language, and most people have not noticed. Every second sentence you read online, in a report, in a marketing email, or increasingly in a novel, was drafted by a machine. The other half was written by a human who almost certainly read something written by a machine yesterday.
We are living through the largest linguistic contamination event in human history, and the interesting question is not whether it is happening. The interesting question is what a language becomes when 50% of it is generated by systems that have no mouth, no childhood, no lover, and no fear of death.
Voltaire once observed that language is a very difficult thing to put into words, which sounds like a joke until you try. He meant that words are not neutral containers. They carry the fingerprints of everyone who ever used them. So what happens when half the fingerprints belong to nobody?
AI and Language: The First Non Human Dialect
Every language in history has been shaped by the physical creatures who spoke it. Vowels shifted because peasants were tired. Slang appeared because teenagers wanted secrets. Grammar simplified because merchants needed to trade with strangers. Language was always the residue of embodied life.
Artificial intelligence introduces something genuinely new: a dialect produced by a system that has never been embodied, never been hungry, never been late for anything. And this dialect is not a small subculture on the edges of the language. It is now a significant fraction of everything written on Earth.
The relationship between AI and language is not the relationship between a tool and its user. A hammer does not change what a builder wants to build. A word processor did not change what writers wanted to say. But a large language model does something different. It suggests the sentence before you have finished thinking of it. It offers the paragraph you were going to write anyway, only smoother, faster, and slightly more generic.
What a Machine Sentence Sounds Like
You already know the tells. The tidy tricolon. The reassuring symmetry. The paragraph that pivots on the word “however” and lands on a moral that no one would ever dispute. The relentless optimism. The absence of odd rhythm, strange metaphor, or the wrong word placed on purpose.
Machine prose has a signature, and the signature is smoothness. It is prose sanded down to the point where nothing catches. There are no splinters, no burrs, no personal grudges. Everything reads as if written by a slightly nervous graduate applying for a job at a company that has no clear function.
When 50% of a language is written this way, the smoothness stops being a style and starts being the language itself. This is the first serious consequence of ai generated text impact: the erosion of texture. And texture, in language, is where meaning actually lives.
How AI Changes Writing at the Level of Thought
Nietzsche wrote that we have to fear that we will never be rid of God because we still believe in grammar. He meant that the structure of our sentences quietly shapes the structure of our thinking. Subject, verb, object. Cause, action, effect. A world of tidy agents doing tidy things to tidy objects.
Now imagine that grammar itself begins to shift, not through the slow pressure of poets and peasants, but through the statistical preferences of a handful of models trained on the internet. What happens to thought when the shapes available to it are chosen by a probability distribution?
A civilization does not lose its language by forgetting words. It loses its language by outsourcing the small decisions that produce style, and style is the last visible trace of individual thought.
Here is the honest answer to how ai changes writing: it does not just change the output. It changes the writer. When a person accepts an AI suggestion, even a small one, they have quietly agreed that the machine’s version was close enough to what they meant. Do this 100 times a day for 5 years and the boundary between what you meant and what the machine offered dissolves.
The Death of the Awkward Sentence
Great writing has always been full of awkward sentences. Not sloppy sentences, but sentences that resist the reader for a moment because the thought inside them is unusual. Read any philosopher worth reading and you will find phrases that make you slow down. That friction is the sound of a mind refusing to say the expected thing.
AI is a machine for producing the expected thing. It is trained, by definition, to predict the next likely word. Its entire purpose is to reduce surprise. This is fine for a customer service reply. It is catastrophic for a culture that depends on new ideas, because new ideas almost always arrive dressed in awkward sentences.
When half of everything written is smoothed toward the expected, the awkward sentence begins to look like an error. Editors reject it. Readers skim past it. Search engines rank it lower. Eventually writers stop producing it, because there is no reward. And with the awkward sentence dies the strange thought that only it could carry.
Vocabulary Compression
Estimates of active English vocabulary vary, but a literate adult uses perhaps 20000 to 35000 words. Large language models technically know far more, yet in practice they draw on a much narrower band. Certain words appear with unusual frequency. “Delve.” “Tapestry.” “Robust.” “Navigate” as a verb applied to anything abstract. “Landscape” applied to any field of activity.
These are not bad words. But when they are produced at industrial scale, they crowd out the alternatives. A young writer reading mostly machine text absorbs machine vocabulary. The pool of words in active cultural circulation contracts, even as the total number of words produced explodes.
This is the paradox of AI language: more text, less linguistic diversity. More sentences, fewer distinct voices. More words per capita, and a smaller usable vocabulary in the culture as a whole.
The Historical Precedent Nobody Talks About
People often reach for the printing press when discussing AI. The printing press is the wrong analogy. The printing press amplified human voices; it did not generate new ones. A better analogy is the introduction of standardized clock time in the 19th century.
Before the railways, every town kept its own time, calibrated to the local sun. When trains started crashing into each other because Bristol thought it was 11 minutes past the hour and London thought it was 3 minutes past, the British government imposed Greenwich Mean Time on the entire country. Within a generation, the felt experience of time itself changed. People stopped consulting the sun and started consulting the station clock.
Something similar is happening with language. Before AI, every writer kept their own local time. Their sentences ran on the internal rhythm of their particular mind. Now there is a Greenwich Mean Prose, and it is being imposed not by government but by convenience, by autocomplete, by the sheer gravitational pull of statistical averages.
The danger is not that machines will write badly. The danger is that they will write acceptably, and that acceptable will become the ceiling of what a human being feels permitted to say.
The Second Order Effect: How Humans Start Sounding Like AI
The most disturbing part of the ai generated text impact is not the machine text itself. It is the human text that starts imitating it. Read student essays from the past 2 years. Read LinkedIn posts. Read internal corporate memos. The rhythms, the transitions, the little rhetorical hedges, all of it increasingly sounds like it came out of a model, even when it did not.
This happens because human beings are extraordinary imitators. We absorb the linguistic environment around us with almost no conscious effort. If half that environment is now machine generated, our own speech will drift toward it. We are not being replaced by the machines. We are becoming them, one autocompleted phrase at a time.
Machiavelli warned that men imitate the paths already trodden by others, and that they should choose great men as their models so that if they fail to reach the same heights, at least something of the greatness will rub off. He never imagined a world in which the most tempting model to imitate would be a probabilistic average of everyone who ever posted a listicle.
What the Great Writers Would Have Feared
Consider what specifically alarmed the philosophers of language across history. Locke worried about the abuse of words, meaning the use of words without clear ideas behind them. He believed that most political and religious conflict came from people wielding words they had never bothered to examine.
Machine text is, by construction, exactly what Locke feared. It is words without ideas. It is fluent output produced by a system that has, by its own architecture, no ideas at all. It has patterns. It has probabilities. It does not have beliefs, doubts, or the small burning need to say one true thing before dying.
Every time a human accepts machine output as their own writing, they are participating in Locke’s nightmare: the circulation of language that no one, at any point in the chain, has taken responsibility for meaning.
The Erosion of Authorial Weight
For most of history, a sentence carried with it the implicit weight of the person who said it. If Montesquieu wrote something about the separation of powers, you could hold him accountable for it. His reputation, his career, his life were on the line.
Machine generated text has no such weight. Nobody stands behind it. And here is the subtle catastrophe: as machine text becomes the majority of text, the expectation that a sentence has an author behind it begins to fade. Readers stop asking. Writers stop caring. The entire epistemic structure that made civilized argument possible, the idea that words are backed by a person who can be questioned, quietly collapses.
The Counter Position: Why This May Not Be a Catastrophe
An honest essay has to consider the strongest case against its own thesis. So here it is.
Language has always been contaminated. Latin destroyed hundreds of local tongues. French administrative language flattened the dialects of France. English is, at this moment, doing the same thing to the linguistic diversity of the planet. Every technology of communication, from writing to printing to broadcasting, has been accused of ruining the language. And yet the language is still here, still capable of surprise, still producing new poets.
Perhaps AI is just the next wave. Perhaps writers will adapt by leaning harder into whatever machines cannot do. If the machine produces smoothness, the human will produce grit. If the machine produces the expected word, the human will produce the strange one. Constraint has always been productive for art. This may be the largest constraint ever imposed, and therefore the largest opportunity.
There is truth in this. But there is also a hidden assumption: that enough human writers will notice the constraint and respond to it. In a world where 90% of published text is machine assisted and the readers cannot tell the difference, the incentive to write differently disappears. You cannot swim against a current that nobody else can see.
The Small Practical Defense
If you write for a living, or if you write because your mind demands it, there are a few defenses worth practicing.
- Read old books. Not because they are wiser, but because they were written before the statistical averaging began. Their sentences have shapes that current models will not reproduce.
- Write the first draft with the AI closed. Not out of purity, but because the suggestions colonize your rhythm before you know it.
- Keep the awkward sentence. If a phrase makes you slow down, and the slowing is because of the thought and not the grammar, protect it from every editor who tries to smooth it.
- Use specific nouns. Machines love abstractions like “landscape,” “framework,” and “ecosystem.” Name the actual thing.
- Notice when your own speech starts sounding trained. Catch yourself using the same transitional phrases. Break the pattern deliberately.
None of this will stop the tide. But it may keep some readable stretch of language alive in your own writing, and that is not nothing.
The Long View: A Language Split in Two
The most likely future is not that AI replaces human language, nor that human language reasserts itself. The most likely future is a split. A high formal register produced overwhelmingly by machines, used for business, education, administration, and most journalism. And a lower, stranger, more fiercely personal register that humans use when they want to signal that a real person is behind the words.
You can already see this bifurcation forming. The rise of handwritten letters as luxury goods. The premium placed on live spoken podcasts over polished writing. The suspicion attached to any prose that sounds too competent. The nostalgia for typos.
Within 20 years, the ability to write a sentence that is unmistakably human, meaning slightly wrong in exactly the right way, will be a valuable and increasingly rare skill. It will be to prose what handmade shoes are to footwear. Most people will not need it. A few will pay a fortune for it.
In a civilization where machines produce the polished sentence for free, the crooked human sentence becomes the only sentence worth reading.
Half the language is now written by machines. That is the fact. What we do with the other half is the only interesting question left, and it is a question that no model, however large, can answer for us. It has to be answered one sentence at a time, by people willing to be strange in public.
Voltaire, who spent his life writing things that could get him arrested, would have understood the assignment. The point of language was never smoothness. The point was to say something that had not been said, in a way that could not be ignored, by a person who was willing to sign their name at the bottom.
That is still available. It is just becoming, day by day, more expensive.
