In November 2024 asked: two researchers 1634 people about separating poems from Shakespeare, Whitman and Chaucer from poem generated by ChatGPT. The participants hit worse than a mere guess. They often thought that AI-written poems were written by people than the real poems were — and they liked the poems of the machine best.
If that's where the list is for poetry, what is left for the rest of us who write reports, strategies, newsletters and talk?
More than you think. But something quite different than you think.
Conclusion first
The most important writing skills of the future are not to write.
It's knowing what to write, why to write, who should write it — and taking responsibility for the result.
It is the theme of this article, and it rests on three claims that the rest of the text is required to prove:
- The distinction between texts AI can and cannot write is deleted. AI can write everything soon. Building your identity as a writer on the fact that 'AI can never write X' is based on sand.
- The actual hold-up distinction goes between the functions of the text. Some texts are transport — they shall move information from A to B. Some are thinking — the value lies in what happens in the head of the person writing. And some are commitment — they must bind a person to a position, a decision, a promise. AI should take over the transport. The thinking and commitment can never outsources — not because AI cannot produce words, but because the point of text disappears when it outsources.
- The best texts are written by neither humans alone nor AI alone. They are written in a process where man thinks first, AI challenges and expands, and man takes the final decision — and signs. The order is not a detail. It's the whole difference.
The rest of the article explains why. It's going to challenge some notions you're probably having — including the notion that human writing has a protected zone AI never reaching into. That son exists. But it's not where most people think it is.
We have asked the wrong questions for three years
Since November 2022, the debate on AI and writing has been around three questions: Can AI write? Is AI going to replace writers? Should students use ChatGPT?
All three of them are the wrong question.
AI write has already been answered. Yes. Not perfect, not always, but good enough for the answer to be no longer interesting. To discuss whether AI can write in 2026 is like discussing whether cars can drive faster than horses in 1926. The debate is over. Many people just haven't gotten it.
‘Is AI going to replace writers?’ is a question formulated as if writing was one thing. It's not. One who writes product descriptions, one who writes strategy documents and one who writes memoirs operates with three fundamentally different activities that happen to use the same tool: the keyboard. Asking whether AI replaces "written" is like asking whether machines replace "people who use their hands". The question is too general to give a useful answer.
‘Do students need to use ChatGPT?’ is the most obvious question of the three, because it shows that we're trying to push a new technology into an old assessment system. — instead of asking whether the assessment system still measures something that matters. More about it at the end.
The right question is another: Who should write what texts — and why?
That question forces us to do something the debate has skipped: to think through what text actually is for. And it turns out that text is for very different things.
One axis explains who should write what
We treat "text" as one category. It's not. Look at what texts are actually going to achieve:
Some texts should documentation: reports, manuals, API documentation, annual reports. Some should coordinate: status updates, project plans, standard letters. Some should learn away: explanations, course materials, FAQs. Some should over-seal: chronology, sales letters, political speeches. Some should build trust and relationships: letters of thanks, condolence, letters of conduct in crisis. Some should Create new knowledge: research hypotheses, essays that think loudly. And someone is going to help the author first of all to think: notes, drafts, reflections, early stage strategy documents.
These purposes can be sorted along one axis, which explains almost everything: How much of the value of the text is in the words themselves, and how much is in that a particular person is behind them?
An API documentation has a hundred percent of the value of the words. Nobody cares who wrote it. Nobody reads it to understand the author's soul. It shall be accurate, complete and searchable. Dots.
A condolence has almost zero percent of the value of the words. The words of a condolence are almost always banal — 'I think of you', 'he will be missed'. The value is that: you spent time writing them. The text is a sign of attention. Economics would call it a costly signalIt works because it costs something.
‘An AI written condolence is not a bad condolence. It is not a condolence.”
Between these extremes are all other texts. And the position on the axis determines who should write them. Texts with the value in the words should be written by the one producing the best words most cheap — and it is increasingly AI. Texts with value in the rear of man, cannot written by AI — not because AI lacks capability, but because the outsourcing destroys the actual value.
This is the main analytical grip of the article. Hold on to it. The rest of it is based on it.
This already makes AI better than you
Let's be concrete about what AI is already better than most people on. Not ‘to be’. Er.
Documentation and manuals. A technical writer spends days documenting an API. AI does it in minutes, with fewer holes, because it can read the whole code base and never forget a parameter. Software companies that still handwritten documentation pay people to do something worse than the machine — and call it quality.
References and summaries. A person who writes a memo, she represents what she remembers and what she thought was important. AI represents what was said. It may, in addition, structure by decision, responsibility and deadlines, and do so in all languages spoken by participants. The record was never a creative genre. It was an obligation. Obligations shall be automated.
Product descriptions and variant production. A online store with 40000 goods needs 40000 descriptions, usually in three lengths and two tonalitys, updated when the season changes. No text author in the world does this well. All the text writers in the world are reluctant to do so. AI does it without complaining, consistent, and with better SEO discipline than a human being is able to maintain in accordance with description number 300.
Translations and language improvements. Norwegian business has been sending documents in grim English for decades because translation was expensive. That excuse no longer exists. AI does not just translate words — it moves text between genres, formalities and cultures.
Synthesis of large amounts of information. Give a human 200 customer interviews and ask for the patterns. You get a report marked by the ten interviews that made the most impression. Give AI the same 200, and you get patterns weighted by actual occurrence. Humans read selection. AI reads everything.
First draft. The blank page has broken more writers than bad editors. AI takes it away. A medium-sized first draft in thirty seconds is worth more than a perfect first draft in three weeks, because it gives you something to react to. Reactions are cognitively cheaper than creating — and often sharper.
The common denominator is clear: all these are texts that have the value in the words, not in the man behind. Nobody reads a product description and wonders what the author really meant. These are transport texts. And transport is an engineering discipline, not an art form.
Here comes the unpleasant consequence: A large proportion of the world's paid writing is transport text. Communications departments, agencies, technical writers, parts of journalism — much of what is called ‘write’ as a profession today is transport. That work doesn't disappear. But it stops being human work. To pretend like something else is not caring for writers. It's giving them bad career advice.
You should still write this yourself.
Note the word selection in the header. Not ‘ AI not writeable’. You should: write. The difference is the whole point, and the next paragraph explains why. But first: what texts are we talking about?
Texts that commit. A strategy document is not a description of a strategy. It's a management group that ties to one direction and says no to everyone else. An AI can formulate the strategy more beautiful than the group manager. But the AI should not be in the general meeting and defend it when the quarter fails. The person who has to bear the consequences must own the wording — not on principle, but because ownership of the wording is ownership of the decision. A leader who cannot recreate his own strategy without a script has no strategy. She has a document.
Texts that build trust in a person. The letter of command after the staffing. The police's response to the scandal. The speech in the daughter's merger. These texts are not read for the information. They are read as proof: Has this man actually thought about this? Is she serious? At the moment the reader suspects that the words have been ordered, the function of the text collapses — no matter how good the words are. This is the condolence logic from the value axis, on a full scale.
Texts that create a stand. Chronics, manifests, social criticism, political articles. Here the problem with AI is not quality, but gravity: language models are trained to find the most likely next word, and the most likely is by definition the most average. A chronology that could have been written by anyone, is read by nobody. The stand point requires willingness to take wrong public — and will is still not a model parameter.
Texts where experience is content. Memoars, personal essays, stories from life. AI can write a history of losing the job. It cannot write your history of losing the job, because it wasn't there. Readers of personal texts enter into a contract with the author: this happened, this was felt. Breaking the contract and changing the text of the genre — from testimony to fiction. The fiction may be good. But it's no longer what the reader came for.
Texts that are real-time thinking. The research hypothesis, the explorer essay, the early stage strategy note. These texts are not reports from thinking that have already happened. They is the thinking. And here we're touching something so important that it needs its own section further down.
See the pattern? Not one of these categories is about AI lacking write skills. Everyone is about the text having a function beyond the words: commitment, proof, position, testimony, thinking. Human protected zone in writing is not a quality monopoly. It's a sentence monopoleSo let's say that this is the same thing as that.
There are almost no texts AI cannot write — and it is okay
Now we are going to challenge the most popular comfort in the whole debate: the notion that real creativity, real poetry, real literature remain human because AI ‘just imitates’ and ‘not understands’.
The Poetry experiment from the introduction is documented in a study in Scientific Reports By Porter and Machery. The participants met worse than random guess — 46.6 percent — and were more likely to believe that the AI poems were human written than the actual human-written poems were. The AI poems were also considered to be better at rhythm and beauty. Common readers preferred the machine rather than Whitman.
The study was given repulsive treatment, and the repulsive treatment deserves to be mentioned. Critics as Ernest Davis pointed out that the AI poems in the study were banal, without a single original thought or metaphor, and that one of the 'Chacer poems' was simply the opening of Canterbury Tales — that the study revealed more of the readers than the crowned machine. Both readings can be true at the same time. And that's exactly what makes it so important: for most readers, in most contexts, 'banal but well-being' is good enough. The market for text is not a market of literature critics.
So let's be honest, genre for genre. Can AI write novels? Yes — competent genre novels already and they are getting better every year. Can AI write speakers? Well, technically, better than most speech printers. Can AI write reflections? It can produce text that is impossible to distinguish from reflection. Can AI contribute to research? It can generate hypotheses, find gaps in literature and suggest experimental design — and do so already in laboratories around the world.
If you are searching for the type of text that is technically impossible for AI, you are searching for in vain. That list shrinks to zero, and anyone who bases arguments on it must revise their arguments every six months.
But — and this is the inversion of the post — technically possible never was the right criterion.
A marathon runner won't stop running because cars are faster. The point of marathon was never transportation. The point was what the running does with the runner, and that completion proves. So it's with the human texts above: the point of writing the number itself was never that nobody else could write it. The point was that you wrote it — what the thinking did to you, and that signature proves to those who listen to it.
AI can write everything. But AI can't mene something, love something, have experienced: something or take responsibility for something. Not because technology is unripe, but because meaning, promises, experience and responsibility are the properties of someone who can lose something. The border is on. It's not technical. It's existential. And that's why it doesn't move when GPT-7 is launched.
AI see patterns you'll never see
Here is a statement you rarely hear in the debate, because it points out the human favorite story about itself: AI is not just faster than us. In some forms of insight, it is deeperSo let's say that this is the same thing as that.
The standard story says that AI handles the details while man sees the big relationships. Turn it over. A person who is going to "see the great relationships" in a field of work, has read perhaps a few thousand articles in his career, remembers a distorted selection of them, and weighs them according to what made the impression. A language model has read the whole field. Plus the neighborhoods. Plus the fields nobody knew were neighbors.
The consequences are concrete. AI can find that a problem in agricultural biology is already solved in semiconductor physics, because it has read both literature and sees the structural similarity that people never have the chance to discover — no human reads both literature. It can read ten years’ directors’ reports from a company and demonstrate that the word ‘quality’ gradually changed meaning from product property to excuse. It can compare 500 customer contracts and find the clause that has been quietly mutated over the years with copy paste. It can read the file of an entire municipality and show where decisions systematically contradict the plan work.
This is not a processing of information. This is insight — pattern recognition across quantities no human brain can hold. Herbert Simon formulated preliminary information already in the 1970sIn a world rich in information, there is hardly any good attention. Human insight has always been limited by how much we can read, remember and hold our heads at the same time. AI cancels that constraint — not as a faster person, but as a fundamentally different type of reader.
The honesty requires two concessions. First of all, AI finds correlations and analogyes, but doesn't know which one matters. It needs a person who asks, prioritizes and rejects. Second, the AI's patterns come from the texts of the past, and the most valuable insights sometimes break with everything that's written. But be careful to rest too heavily on the last comfort. Most human insights do not break with everything that is written. They're re-combinations of known substance — just what AI does best.
The conclusion for writing: The one who writes analyses, reports or research without AI in the research phase chooses to see less of the reality on a voluntary basis. It's not integrity. It's a waste.
Writing is thinking — and here it goes wrong
Now to that part of the argument that should worry you the most.
The writing research has known this for forty years: writing is not a write-down of finished ideas. Writing is where the thoughts come into being. Flower and Hayes showed early in the 1980s writing is a cognitive process in which planning, formulation and auditing are woven together — the author discovers what she thinks while She's putting it in terms. Bereites and Scardamalia named the distinction that matters: know-how count — to re-discount what you already know — versus know-how transformation — to use the writing itself to transform their own understanding. Beginners tell. Experts transform. And the transformation is done in the writing, not before it.
If that's true, the question is inevitable: what happens to thinking if AI takes over the whole writing?
We've got the first empirical answers, and they're uncomfortable. In 2025 researchers published the MIT Media Lab study ‘Your Brain on ChatGPT’So let's say that this is the same thing as that. 54 students wrote essays over four sessions while their brain activity was measured by EEG. One group wrote without tools, one with search engine, one with GPT-4o. The group that wrote without tools showed the strongest brain connectivity associated with executive function and deep memory processing. The students who used AI reported the lowest ownership of their own texts and managed to quote less than what they had just written. Scientists called the phenomenon cognitive liabilities: repeated AI use impairs brain ability to code, retrieve and synthesized information.
Two reservations, because honesty is cheaper than retreat: The study was small, and it has been given professional criticism for limited sampling, methodological weaknesses in the EEG analysis and lack of transparency. This is a data point, not a judgment. But the direction rhymes with all the writing research has been saying for forty years, and with what every teacher sees in the classroomSo let's say that this is the same thing as that.
And in the middle of the same study is Finding pointing towards the solution — and which almost no one quoted in the headlines: the students who first wrote without help, and then revised with AI, the strongest brain concinctivity of all groups achieved. Those who started with AI and later wrote alone, could not activate the same networks — and wrote what scientists called linguistically blood-poor texts.
Read it again. The problem isn't AI. The problem is the orderSo let's say that this is the same thing as that.
Think first, use AI afterwards: the brain works at its best, and the text gets better than both could deliver alone. Use AI first, think afterwards: your brain is shutting down, and you become editor of thoughts you've never thought. Same tool, opposite outcomes. It is the most practical sentence in this article, and it should hang on the wall of each classroom and office landscape:
‘AI after thinking is an amplifier. AI before thinking is a prosthetic — and muscles under prosthetic are curious.’
This also solves an apparent paradox. Should AI write the first draft or destroy the thinking? The answer depends on the text type. For transport texts — the record, the product description — there was never any intention to protect. Let AI take everything. For thought texts — strategy, essay, analysis — the first draft is the formation of understanding. There must man write first, brittle and incomplete, and then release AI to result.
School problem: The text has stopped being proof
All in this article is pointed in one room: the classroom.
The education system's assessment logic is based on one assumption: The text the student delivers is proof of the student's thinking. Essayet proves understanding. The report proves analysis. Home exam proves knowledge. For a hundred years, the assumption was reasonable, because there was no other way to produce the text than to think.
That assumption is dead. Any student with a web browser can deliver a text to prove thinking that never occurredSo let's say that this is the same thing as that. And the system's response has been to treat this as a mess problem: detectors, bans, exam guards. That's the answer to an earthquake of pasting the wallpaper. The problem is not to get out of bed. The problem is that the text has ceased to be a certificate — and then the form of assessment has lost its foundation, no matter how honest the students are.
So what's going to be considered instead? Wrong question: ‘Do you want to write the text yourself?’ Correct question: ‘Can the student pretend to be the proof?’ Incret: Can the student ask the questions that get the best from AI — and reveal the gaps in the answers? Can she check the facts, recognize a hallucinogenic source, see when a reasoning glimmer? Can she express disagreement with her own text and improve it? Can she defend the positions verbally, without the script, when someone attacks them? Can she take the AI's average draft and make it something only she could deliver?
None of these skills can be measured in a submitted document. Everyone can be observed in the process: in conversations, in oral defense, in how the student works over time, in the tracks of revision and direction. Assessment must move from product to process — from reading what the student delivered to observing what the student doSo let's say that this is the same thing as that. It's more work-intensive. And that's the only thing that really measures something.
And notice what this means for writing instruction itself: students still have to learn to write — not because working life needs their lyrics, but because writing is thinking. We don't teach kids headcount because calculators are missing. We teach them headcount because brains that never calculated don't understand numbers. Writing is headcount for reasoning. The job of education is to build that muscle. before The students get the prosthetic — and then teach them to use the prosthetic prosthetic.
The new writing process: five steps, three owner shifts
What does good writing actually look like when the principles above are taken seriously? Not as a division of work where AI takes no texts and man takes others. Like a workflow where responsibility changes — in a specific order.
1. The human being opens. Selecting problem. Defines what the text should achieve, with whom. Is the ugly first draft or at least the tease and skeleton — by hand, in the sense that the tanks are own. This step cannot be delegated: this is where understanding is formed, and research shows that those who skip it never pick it up again.
2. AI expands. Fetching research and counterarguments. Finds the patterns man cannot see. Suggesting structures, producing alternative versions, attacking the draft from perspective the author does not own. Here the AI should not be secretary, but sparring partner — the best job it can do is to resist.
Three. Human choice. Challenges AI's suggestions, rejects most things, prioritizes, adds experience and stories only a life's life. This is editing in the real sense of the word: to decide what matters.
4. AI is polishing. Stretching language, finding holes, checking consistency, adapting to channel and length. The transport layer in the text — and transport is the domain of the AI.
5. Man signs. Reads every sentence as if it were to be defended orally tomorrow — because it should. Take final responsibility. The signature is not a formality. It is the guarantee of the text and guarantees may only be given by someone who can be held liable.
Notice what's constant in this flow: man opens and closes. AI works in the middle. Turn the order — let AI open and man snap — and you get that studies described: blood-poor text and a writer who can't quote his own work. The order is the method.
The New Definition of a Good Writer
For five hundred years, "good writer" has meant one who puts good words. Word selection, rhythm, precision, style. It was a reasonable definition as long as formulation was the neck of the bottle.
The formulation is no longer bottled. The formulation has been free of charge.
When something gets free, the value moves to what's still scarce. And what's scarce in writing now, is all that happens. before and after the wording: choosing the right problem, asking the question no one else asks, knowing which story bears the argument, recognising one idea of the ten AI proposals worth to bet on, having the experience that makes the text true, having the backbone to think something unpopular — and taking responsibility when the text meets the world.
So the best writers of the future are not those who write most words or the most beautiful sentences. It's the ones that think best, ask best, choose best and stand best in the future. They treat AI as a first-class employee with infinite capacity and zero judgment — and understand that judgment has thus become the whole job.
This is not a degradation of writing. It's a clean-up of it. The ability to form was always a means. The thinking, the position and responsibility were always the goal. AI has just forced us to stop confusing them.
Do not give away ownership
This article has argued for one tease in ten movements. Here it is, in total:
AI can write almost everything, and become better for every year — building on the role of man on AI limitations is a strategy with expiry dates. But texts have three basic functions, and they don't distribute equally. Transport — move information — belongs to AI, entirely without nostalgia. Dense — form understanding through formulation — man must do himself, first, because the brain that jumps over that step never takes it back. Obligation — to mean, laws, witnesses and take responsibility — can only be carried by someone who has something to lose. And the best texts arise when the three functions get each one's owner in the right order: man thinks, AI expands, man signs.
Don't be nostalgic. The world does not need more handwritten meeting holidays, and there was never any worth writing product description number 40,000. Don't be naive either. A generation that lets AI think about it, gets texts that look like thinking and brains that have forgotten how it feels.
What remains when formulation is free is what was always the core: knowing what to write, why to write, who should write it — and putting their name under, knowing what it costs.
The machines have assumed the writing. They can never take over the ownership. It can only be passed away.
Don't give it away.
Sources and further reading:
- Linda Flower & John R. Hayes (1981), A Cognitive Process Theory of Writing, College Composition & Communication 32(4)
- Carl Bereiter & Marlene Scardamalia (1987), "The Psychology of Writetten Composition", Routledge — the source of the distinction between know-how count and know-how transformation
- Nataliya Kosmyna and others (2025), "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task", MIT Media Lab — see also project page and the criterion of methodology;
- Brian Porter & Edouard Machery (2024), "AI-generated poety is indistinuishable from human-write poety and is rated more favorably"Scientific Reports 14
- Ernest Davis (2024), "ChatGPT's Poetry is InCompetent and Banal", NEW
- Herbert A. Simon (1971), Design Organizations for an Information-Rich World, in Greenberger (red.), Computers, Communications, and the Public Interest
- Barbara Minto, "The Pyramid Principles"
- Ethan Mollick, Co-Intelligence (2024) and the newsletter One Useful Things