AI Writing Is a Fate

Author: Wang Huanchao, Senior Researcher, Tencent Research Institute

In 1882, Friedrich Nietzsche — thirty-seven years old — bought a fashionable machine. It was called the Malling-Hansen Writing Ball, designed by a Danish pastor for the blind: fifty-two keys arranged in a hemisphere, like a steel hedgehog.

Something like this.

Of course, Nietzsche didn't buy it to be trendy. His eyesight was failing. Staring at paper for long stretches made his head ache violently; he could barely manage extended handwriting anymore. He needed a way to write that didn't depend so much on vision.

After Nietzsche had used the writing ball for a while, his friend, the composer Heinrich Köselitz, wrote in a letter: "Perhaps with the aid of this new instrument, you will obtain a new mode of expression." Köselitz's own thoughts in music and language, he said, often depended on the quality of pen and paper.

Nietzsche agreed. His reply contains a line quoted ever since: "You are right — our writing tools are also working on our thoughts."

Later researchers noticed that the prose Nietzsche produced on the writing ball was more compressed than his already aphoristic style. The long, winding, clause-stacked sentences disappeared, replaced by shorter ones — almost telegraphic.

A century later, German media theorist Friedrich Kittler, in Gramophone, Film, Typewriter (1986), elevated this episode into a key testimony of his media theory. In Kittler's reading, Nietzsche's failing eyes are a metaphor. Writers do not first have thoughts and then pick a convenient tool to express them. On the contrary: the tool first delimits a region in which thought can take shape, and thought grows form inside that region.

Change the tool, and you change the way of thinking.

I want to start from here to talk about AI writing.

Today, nearly all discussion of AI writing stays on one plane: to use or not to use. Is using it cheating? Is it still "yours" if you use it? Do you have to disclose it? These questions are legitimate, but they share a suspicious premise — that we are standing at a fork in the road, and still holding the choice in our hands.

My judgment is that this fork does not exist.

AI writing is not an option. It is a fate.

Of course, "fate" here is not literary embellishment, nor a fatalistic surrender. It is closer to what Heidegger called Geworfenheit (thrownness): you don't pick one among several possibilities — you open your eyes and you are already thrown into a situation.

Nietzsche could, of course, have declined to buy the writing ball. The cost was to stop writing. That is what a fate means: you seem to have choices, but all of them sit on the same side.

The genealogy of the fear

Look across the history of media tools, and you find that resistance to new fears is cyclical.

It begins with Plato's Phaedrus. Socrates tells an Egyptian story: the god Theuth invented writing and presented it to King Thamus as a medicine that would make Egyptians wiser and improve their memory. Thamus refused. Writing, he said, would produce not memory (mnemé) but reminders (hypomnesis). People would trust external symbols and cease to call up knowledge from within. They would seem learned while being only "seemingly" so.

This is humanity's first full articulation of the fear of cognitive outsourcing — and in hindsight, its accuracy cannot be denied.

The list goes on from there. The printing press was accused of making people "read without thinking." The typewriter was condemned by Heidegger for "snatching writing from the essential realm of the hand," reducing words from traces of being to mere means of communication. Word processors were blamed by critics for erasing the traces of revision. As for the internet, it became a lightning rod — Nicholas Carr's The Shallows warned it would cost us deep reading and continuity of thought.

Every media transition, every new tool, triggers a wave of fear and criticism.

Yet notice: in the history of technological outsourcing, only cognitive outsourcing provokes ethical critique. Physical outsourcing almost never does. Nobody criticizes cranes, steam engines, or washing machines — the occasional Luddite is quickly drowned in the current of history. But every time outsourcing touches our heads — memory, writing, retrieval, expression — the alarm sounds on schedule, and critique follows. Presumably because physical outsourcing only uses muscles and extends the body, while cognitive outsourcing concerns "who I am."

AI clearly sits on the extended line of this genealogy of fear.

But this time, the fear does contain something genuinely new. In all previous cognitive outsourcing, the boundary stayed at external storage: writing performed memory, print performed dissemination, search engines only retrieved, word processors helped you revise. The hardest and most central step of writing — compressing chaotic intention into a sentence, compressing a web of thought into linear prose — still had to be done by a human.

Generative AI is the first technology to outsource that very act of compression.

So this round of panic has legitimacy. What is illegitimate about it is the direction of the fear. Most people are still agonizing over "should I let it write for me," while what has actually happened is that even if you never let it generate or output anything, every element you rely on to write has already been swapped out from under you. That is why I call it a fate.

There is no outside

A basic economic principle: when a tool pushes the cost of some step toward zero, those who don't use it must bear the cost difference alone.

Of course, a cost difference can be filled with willpower. Today there are still writers who insist on handwriting first drafts, and photographers who shoot only film — they spend more time and money to buy a working method they recognize as their own. As long as the price is bearable, exit is possible. If the non-exitability of AI writing stopped there, it would be a mere choice, not a fate.

In 1970, economist Albert Hirschman, in Exit, Voice, and Loyalty, gave "exit" a plain definition: a dissatisfied customer turns and buys another company's product. But this definition has a precondition — exit requires a place to exit to. Only if an outside exists does exit hold and make sense.

The problem now is that the outside one could retreat to is disappearing, layer by layer. There is no outside anymore. Everyone is inside the situation, and no one can exit.

In July 2025, Dmitry Kobak's team at the University of Tübingen published research in Science Advances. They compared more than 15 million PubMed abstracts from 2010 to 2024, hunting for "excess vocabulary" — words that suddenly exceeded historical trends in 2024. What exploded, they found, was a set of style words: "delves," "showcasing," "underscores," and the like.

By contrast, over the previous decade the words that grew were mostly content words — "Ebola," "COVID" — corresponding to things happening in the world. In 2024, what surged were almost all verbs and adjectives, corresponding only to a way of speaking. Based on these traces, the researchers' lower-bound estimate: at least 13.5% of biomedical paper abstracts in 2024 were processed by large language models.

The significance of this figure lies in its form. The first consequence of AI's transformation of writing to be measured at scale is not content fabrication, content padding, or "AI slop" — it is stylistic convergence: more and more people speaking with the same vocabulary.

In June 2025, Nataliya Kosmyna's team at the MIT Media Lab ran an EEG experiment. Fifty-four participants wrote argumentative essays in three groups: pure mental effort, search-engine assisted, and LLM-assisted. The researchers ran named-entity and n-gram analyses. Their conclusion: the LLM group's essays converged heavily in topic and wording. Given the same prompt, the pure-mental group produced all kinds of angles; the LLM group produced variants of the same essay.

Two other results matter. First, the LLM group showed the weakest neural connectivity in relevant brain regions. Second, shortly after finishing, 83% of the LLM group could not accurately cite the sentences they had just "written." The researchers named this state "cognitive debt."

These mutually similar texts — which even their nominal authors cannot remember — are flowing continuously into the public space.

Even if you never let AI touch a single word, your readers, your editors, your reviewers, your recruiters read large volumes of AI-assisted text every day. Their standards for what counts as "fluent," "well-structured," "a competent opening," or "moving expression" are being reset by these texts.

As more and more people's prose converges toward the same median, that median becomes the new standard.

Once a standard exists, deviations from it are judged "not fluent enough," "not regular enough." Those who insist on not using AI are not merely "odd" — they may be ruled simply unqualified.

More crucially, you yourself live inside this cultural ecology. When everything in the ecology is shifting around this standard, and you absorb it daily through immersion, what grounds do you have to claim you have not been changed?

This is not only a rhetorical question. In 2024, Hiromu Yakura, Iyad Rahwan and colleagues at the Max Planck Institute for Human Development analyzed more than 360,000 YouTube academic talks and 770,000 podcast episodes around the release of ChatGPT. It was a massive undertaking — 740,000 hours of spoken material — to test whether words identified as ChatGPT-preferred would enter human speech.

The answer: yes.

After ChatGPT's release, "delve," "meticulous," "comprehend," "swift," "boast" appeared markedly more often in people's spontaneous speech.

The greatest significance of this study lies in its material. Academic papers might plausibly be handed to a model for polishing, but improvised talks and casual podcast chat can almost never be generated verbatim by a model. These words were not written by AI — they were spoken by people themselves. They entered human linguistic habits directly.

This is common knowledge about language acquisition that we rarely think of alongside AI. No one "chose" their mother tongue. A child learns to speak by immersion — they learn what they hear. Adult language never stops growing either; it just grows more slowly, and more invisibly.

The text you read most quietly becomes the text you write most readily.

At this point, one last retreat seems to remain: conscious resistance. Since AI has preferred words and sentence patterns, why not deliberately avoid them?

That path fails too.

In 2023, James Zou's team at Stanford published a study in Patterns. They fed 91 TOEFL essays from a Chinese education forum to seven mainstream AI detectors. On average, over 60% were flagged as AI-generated. The reason: non-native writers have limited vocabularies, regular sentence structures, and common word choices — which are precisely the statistical features detectors use to identify "machine text."

This study is often cited to argue AI detection is unreliable. But it also reveals something else: the standard for whether a text sounds human has already been defined, in reverse, by machines.

Writing plainly, regularly, and simply now risks suspicion of being AI-generated. If you want to be confirmed human, you must write unlike a machine — and what "unlike a machine" looks like is again defined by machine standards. The more diligently you try to dodge the machine's norms, the more your dodge is already measured against them.

The intended writer who exits, who resists, is forced into the position of being defined by the very thing he resists.

So in which layer does the boundary really sit?

What AI takes, and what it gives

Thus far this essay has largely narrated coercion. But if AI writing is a fate, it is not a one-sided fate.

For two hundred years, every wave of technology that made a scarce skill cheap also made something once common newly scarce. Print made handwriting cheap — and manuscripts became treasures. Photography made portraiture cheap — and painting broke free from the duty of realism. Generative AI makes the skill of turning internal states into readable text cheap — and what becomes precious is precisely what cannot be generated: experience and judgment, the moments of thinking that precede compression.

In 2006, Tencent Research Institute's director, Xing Bo, wrote a well-known short text, the second originals: true originality takes two paths — saying something never said before, or saying the same thing in a meaningfully different way.

This distinction is extremely useful for what follows.

Let me first ask the question that hangs over every discussion: will AI kill writing, will human writing wither?

My answer is no.

What is dying is the instrumental use of writing — writing that serves only as a means to an end: the commodity article, the formulaic report, the standardized official document, the routine copy. These are the median texts that AI already produces more cheaply, faster, and more voluminously than people.

But writing as such has never been merely instrumental. What generative AI cannot do is be responsible. A text written by AI, when asked "what did you mean by this," answers with majestic silence. It cannot be questioned, cannot be hurt, cannot change its mind, cannot bear consequences for a single sentence. Writing that can be questioned, can be hurt, can change someone's views, and can bear consequences — this kind of writing subject becomes scarcer the more text becomes abundant.

What is scarce, in the end, is the writer's presence.

Literature after AI

Now consider literature — in the broadest sense, the strongest writing tradition.

If the essence of literary value lies in language, then AI, trained on the largest corpus of human text, is precisely the most "literary" machine ever. No human can outperform it on conventionalized language. Under the old framework — talent is the rare possession of a few — writing literature is already, to a large extent, a competition with machines.

But there is a deep misunderstanding in this comparison: literature was never won by language alone.

In 2024, researchers at the Chinese University of Hong Kong, Shenzhen compared human-authored novels with works written by LLMs. On seven dimensions including plot design, character creation and narrative structure, the AI novels did not lose — on some metrics they even scored higher. Yet human readers still overwhelmingly preferred the human novels. When asked why, they could not articulate it clearly. They just said: the AI novels, no matter how they are tuned, feel like they are "missing something."

What is missing is not language but weight — the weight of lived experience that language carries. Language can be generated, but weight cannot.

The distinction is roughly that between having a language and having something to say. The former is a skill; the latter is a destiny. AI gives everyone the skill; what it cannot give is the destiny.

This is also why I believe literature will not die but will be reassigned to a more precise position.

Just as the dictionary does not kill the poet, and the camera does not kill the painter. On the contrary, the more abundant the means, the more the scarce end is worth — and the more fiercely those who choose it are driven to the extreme of their own end.

The history of literary genres is not a graveyard of extinct forms; it is a chronicle of value being relocated. Writing platforms, genre conventions, readership patterns all shifted, but the people who insisted on writing with their own blood never disappeared. They just moved to the next margin.

Freedom begins with the admission of fate

There is an older and stranger question behind the debate about AI writing — not "can AI write well," but "who deserves to write."

For most of history, this threshold was guarded by education, literacy and class. The ability to turn thoughts into readable text belonged to a few. Most people's experience and judgment never became text — not because they were worthless, but because the means to express them were controlled.

In a long stretch of Chinese history, there were letter-writing stalls at post offices and railway stations. Those who came were mostly people who could not read or write well. The hired scribe wrote the formulas — "Honorable parents above," "reading this letter is like seeing my face" — which were common property. But the "individual content" of the letter — this year's harvest, a child's illness, the hardships and longings of migrant work — could only be spoken by the person dictating it.

Writers like Fan Yusu, Yu Xiuhua, Hu Anyan and Wang Jibing are rare not because such experience is rare, but because people who can turn experience into fluent text are rare.

Generative AI is the first technology to lower this threshold to a height almost anyone can cross. A man who has worked thirty years on a construction site, a researcher whose first language is not English, a student with dyslexia, a person full of ideas but unpracticed in writing technique — all can use AI to turn what is in their heads into text others can read.

This is technological democratization. What is more telling: what AI gives everyone is precisely what it is best at; what these people already have is precisely what it most lacks — experience and judgment.

AI supplements the ability to express, not the content of expression.

Combined, the two could give birth to countless texts that never had a chance to be written.

But there is a precondition. Democratization and convergence are two faces of the same thing. The construction worker who uses AI to write his thirty years — that is democratization. But if he hands over what to write and how to judge as well, then what he writes is only the valueless median.

AI gives more people the ability to write. But only those who do not fully submit to it truly own their words.

Back to the question at the start of this essay. "To use or not to use" is a wrong question because it presupposes an outside one can exit to. Since the outside no longer exists, the question that actually needs answering is only one: at which layer to use.

This is a migration of ethical coordinates — from an ethics of choice (use or not) to an ethics of division of labor (what did I hand over, and what did I keep). And it determines that our way of writing will certainly change a great deal.

First: draw the load-bearing walls, then build. Order matters. The most common AI workflow today is to let the model generate the full text first, then have the human revise and polish. This order is easily lost. When a complete, coherent, clearly argued draft sits before you, your judgment has already been locked into its framework; it is very hard to break out of the frame it has set. Like the Sophons locking human technology in The Three-Body Problem, AI's sixty-point draft locks our structure of thought. The correct order is the reverse: first write those few load-bearing sentences yourself — a judgment, a piece of personal experience, a position — then hand the expansion of structure, transitions and examples to the model.

Second: deliberately preserve a little inefficiency. Stylistic convergence comes from the accumulation of local improvements: swapping in a more common word, adding a smoother transition, evening out irregular sentence lengths. Each one, taken alone, is an improvement; accumulated, it is a convergence toward the median. The way to resist is to actively keep some places unimproved. This is not manufacturing errors, broken grammar or clumsy phrasing. To deliberately write haltingly in order to seem human, deliberately unlike a machine, is still to be defined by the machine. What you should keep is what is already yours: a sentence that isn't smooth but only you would say it that way; a slightly overlong aside that carries the rhythm of your thinking; an uncommon usage you insist on. Deliberately preserved inefficiency is the recognizable signature of an individual.

Third: put your energy into the second kind of originality. In Xing Bo's essay, originality has only two paths: say something never said, or say the same thing in a meaningfully different way. The claim traditionally attributed to the Pythagorean school — that the earth is spherical — belongs to the first kind. About a century and a half later, Aristotle said the same thing in On the Heavens, but he supplied empirical evidence: the round shadow cast on the moon during eclipses, the changing stars one sees traveling north and south. That is the second kind. This distinction matters now because it dissolves a false pressure. Most writers will never produce the first kind of originality in a lifetime — that is the norm. The value of the second kind is, on the contrary, raised by AI: when commodity content is supplied without limit, retelling an old thing or common knowledge in a way that only you could, with your "individual content," becomes increasingly scarce.

Socrates' final warning

In Phaedrus, King Thamus' warning about writing is famous. But after telling that story, Socrates adds a second warning in his own words, far less known.

Writing, he says, is like painting. The figures in a painting stand lifelike, but ask them a question and they preserve a solemn silence. Writing is the same: you consult it and it does not respond; it forever repeats only the same sentence. The tragedy of writing is that once it is written down, it circulates everywhere — reaching those who understand it and those who do not. It does not know to whom it should speak and before whom it should fall silent. When misunderstood and attacked, it always needs its "father" to come to its rescue, for it cannot defend itself.

That "father" is the author. All the fragility of writing comes from the author's absence.

In the age of print, the author's absence was a defect of writing — but behind every text, at least one real person had once stood. In the age of AI, vast volumes of circulating text may have no one standing behind them who was ever present.

If AI writing is a fate, the final destination of that fate can be condensed into one sentence: the supply of text becomes infinite, the presence of authors becomes scarce.

This is also why I do not believe writing will die or decline. Every time technology makes some capability cheap and infinite, its opposite is redefined and re-priced. Print made copying cheap — and manuscripts became collectibles. Photography made portraiture cheap — and painting broke free of the obligation to be realistic. Generative AI makes words cheap — and the writing subject who can be questioned, can be hurt, can change views, and can bear consequences for a sentence becomes, for the first time, a premium.

In 1882 — the same year the writing ball arrived beside Nietzsche — he completed The Gay Science. At the opening of its fourth book, he made a wish for the new year: "I want to learn more and more to see as beautiful what is necessary in things... amor fati — let that be my love from now on!"

You may read "love of fate" as resignation. Or you may read it differently: stop spending your strength pretending fate does not exist, and keep your strength for the things truly decided by you.

AI writing is a fate. To acknowledge the existence and intransigence of fate has always been the beginning of freedom.