Let's cover the logos on three presentations. One concerns selling flats, another a new app, the third a restaurant strategy. The colours differ, but each arranges the world in much the same way: a problem, an opportunity, three benefits, and a future we're prepared for. It all looks reasonable. Only after a moment does the suspicion arrive that these companies share an author, though they have probably never heard of one another.
With articles and reports produced with AI's help, I sometimes find myself searching for something that would hold my attention on this particular text. The sentences are correct, the paragraphs obedient, no thread left without a courteous closing. "You can feel the machine" would be a convenient explanation. Except that people were capable of writing this way long before language models appeared. Who commissions such texts, and what exactly do they consider good in them?
I don't stand outside this problem. I asked an AI for an essay about the human voice, written so as to sound human, and then commissioned a second model to review it harshly. The response included the charge that an article warning against regular punchlines was itself producing them regularly. That's a rather awkward starting point for reflections on one's own voice. I'll return to that review, because alongside an accurate diagnosis it contained a peculiar course of treatment.
Smoothness on Demand
In corporate language, a delay can become a challenge, an objection a need for further clarification, and a person a stakeholder. Such substitution has its uses: it allows a document to be passed onward without opening a dispute that would require a decision. Imagine an employee who knows their manager accepts only materials that sound "professional." The employee doesn't need instructions on concealing doubts. It's enough that they see a few times which version comes back for revision and which makes it to the board meeting.
AI can accelerate this adaptation. For the author, the saving is editing time; for the manager, the familiar order of argument. The individual document becomes more efficient, while the question that was missing from it remains invisible. The cost of an unused idea is harder to see, because it appears on no slide.
Paul DiMaggio and Walter Powell described organisational convergence as early as 1983. Under conditions of uncertainty, firms imitate solutions deemed credible; shared professional environments propagate norms of what counts as appropriate. Adopting a template can bring acceptance even when its practical advantage has not been demonstrated. [1] Reading them today, I think of a presentation that looks the way a presentation is supposed to look. Someone must once have established that strategy fits neatly into three arrows.
This is a sociological frame for my suspicion, not a technical description of a model. Textual similarity may also stem from how the tool works, from similar prompts, or from the requirements of the genre itself. What interests me, however, is the recipient's share in it: if we approve the same kind of argument every time, then a tool that saves labour on producing it arrives at a well-prepared market. Smoothness is sometimes an order placed. Humanisation is then meant to add a touch of temperament to a statement whose fundamental order no one is permitted to disturb.
One's own voice is also easily mistaken for ornament. A short sentence, a joke, or rough syntax can be generated just as competently as a corporate paragraph. I learn far more about an author when I see what they considered important and what they're prepared to argue about. But a model can propose that point of view too. If we want to establish where the human sits in this collaboration, the sound of the sentences alone quickly ceases to suffice.
A Reader Without a Label
The reader may not share my weariness at all. In an experiment by Brian Porter and Edouard Machery, people identifying the origin of poems achieved 46.6 percent accuracy — slightly below chance. They more often took ChatGPT-generated works for human ones than they did actual poems by poets. In a separate part of the study, AI poetry was rated more favourably on many dimensions. [2] That's an inconvenient result for anyone who believes a reader needs only a moment's concentration to feel the difference.
Short works in English were studied, mainly among people who rarely read poetry. Self-declared familiarity with poetry or knowledge of a given poet was not associated with better recognition. [2] Under other conditions, recognition may succeed: Jason Chein and co-authors recorded above-chance results, depending among other things on the type of task. [3] We therefore have no single verdict on creative work as a whole. What we do have is sufficient reason to stop treating our own conviction that "I would know" as a special literary qualification.
It's easy here to make a defensive move and declare that the participants simply didn't understand good poetry. Except that their enjoyment of the reading was part of the phenomenon under study, not an error to be removed from the results. We read many texts in order to understand something, to feel something, or to get something done. When a model does this well, its origin does not invalidate the effect. "Sounds human" is, besides, modest praise. Humans can sound all sorts of ways — including like car park regulations.
From among three presentations, then, I won't choose the best simply because its sentences were written by someone at a keyboard. I may rate a document prepared with AI more highly if it explains the problem better and can be verified. The wish to defend a human author should not compel anyone to defend their weaker text. Only after acknowledging this possibility can one sensibly ask whether something is nonetheless lost in the widespread use of such assistance.
A Better Text in a Similar World
In an experiment by Anil Doshi and Oliver Hauser, participants wrote eight-sentence stories. Some received ideas prepared by GPT-4. The assisted texts received better ratings on average, particularly when their authors had scored lower on an initial creativity test. At the same time, the stories became more similar to one another. [4] Both things happened simultaneously. The benefit to an individual author was not a good measure of what was happening to the collection as a whole.
Let's transfer this problem, now as a hypothesis, into a company. Every department wants to submit better material. Every one reaches for convenient help. None is responsible for the diversity of ideas in other departments' documents, and certainly none receives a bonus for the fact that a different company might think differently. In such a configuration, similarity can accumulate without anyone intending to standardise anything. The warning "use AI carefully," addressed to an individual employee, covers only a small part of this mechanism.
The shift can begin even later, when the idea is already one's own and the model is merely meant to improve the style. Zhivar Sourati and co-authors found, in the English-language corpora they studied, that such rewriting reduced the variation in linguistic complexity. [5] Editing form can therefore erase differences the author never meant to remove. The measured features of language are not a full portrait of a personality, but neither does the word "smoothing" guarantee the innocence of the procedure. A shorter sentence may have cut a thought short; it may equally have marked its boundary precisely.
The inevitability of this process would, however, be too strong a conclusion. Yun Wan and Yoram Kalman used more diverse inspirations in a small experiment, prepared by AI on the basis of fictional profiles. They did not reproduce the earlier homogenisation effect in the stories; some results suggested an increase in diversity. [6] The study altered several elements of the procedure, so it offers no simple recipe for a remedy. It does suggest that how inspiration is delivered may matter.
This leaves us with a problem of tool design and of the rules governing their use in organisations. A careful author can work out something distinct. They have no access, however, to all the other texts with which their work will one day form a shared landscape. Even if everyone sensibly improves their own document, the diversity of the whole will not improve automatically as a result.
Help Comes with an Accent
A demand for greater self-reliance sounds different coming from someone for whom writing comes easily. In a company, one can know a production process superbly and be unable to describe it in a way that gets the report taken seriously. One can have an accurate observation but be searching for words in a second language. Such a person doesn't need another essay on the nobility of wrestling with a sentence. They need a way for their knowledge to be heard at all.
In a study by Shakked Noy and Whitney Zhang, professionals using ChatGPT completed specified writing tasks in 40 percent less time on average, and the quality ratings of their work rose by 18 percent. Those with weaker initial scores gained the most. [7] It's easy to see in these figures nothing but saved hours. For a person whose ideas had until now been lost in clumsy form, an improved text can change their position in the conversation.
The trouble is that linguistic help arrives with some conception of correct expression. A study of people from India and the United States writing in English showed that AI suggestions homogenised the texts, and that the writing of Indian participants shifted toward American patterns. [8] Gaining easier access to the conversation may therefore mean consenting to someone else's way of telling a story. This is not proof that every form of help in a second language strips away distinctiveness. It does show a cost that assessing the fluency of sentences alone may not reveal.
These two studies do not add up to a simple ledger in which benefit offsets loss. They concern different people and different tasks. They set two needs side by side: the ability to speak up and the ability to preserve the distinctiveness of that voice. Is an organisation prepared to recognise both? If equal opportunity consists in admitting everyone on condition that they speak like those already admitted, the tool may perform its task brilliantly while the rule remains unchanged.
In evacuation instructions, shared vocabulary is a virtue. In departmental reports, a similar structure makes data easier to compare. Distinctiveness need not be defended in every line. It is worth knowing, however, when order helps us see a difference and when it makes everything look just as it did before. This is precisely where an editorial decision meets a decision about what information a company wishes to hear.
Two Dry Jokes
The reviewer of my essay had a strong argument: a text about one's own voice too often ended paragraphs with sentences that looked ready to be underlined. It also noted that caution toward sources had begun to substitute for developing the thought. Then it proposed a cure for monotony: two long sentences, one digression, and two dry jokes in the second half. The diagnosis concerned mechanicalness. The prescription allowed it to be carefully planned.
This small episode is more interesting than the assurance that a machine doesn't understand humour. The model identified a problem that earlier work with AI had not removed, and justified it with specific passages. It helped in critiquing the text. At the same time, its recommendation could be executed without pausing to consider where, in this particular essay, a digression was supposed to come from, or what a longer sentence would be needed for. One could commission the revision and tick off every ingredient.
This is precisely why humanisation is such an imprecise order. A model may read it as a task of supplying signs of a human being: a slightly uneven rhythm, a personal tone, an aptly chosen detail. All of these can work. In a story, an invented small detail can be superb; in a report about a company, it should have corroboration. Here, the sound of a text, its value, and its origin begin to diverge. The most convincing voice need not belong to someone who lived through the scene described.
Origin matters nonetheless, even when the words themselves are good. In experiments concerning personal experiences, AI responses gave participants a stronger self-reported sense of being heard than responses from untrained strangers. Being informed that a message came from AI weakened this effect. With accurate labelling of authorship, the two influences roughly balanced out. [9] Short exchanges were studied, not friendship or therapy. The result is sufficient, however, to show that a recipient also responds to whom they attribute the attention to.
A letter from someone close can be treasured for sentences an anonymous judge would find clumsy. When preparing instructions, that same person may prefer a model's more efficient text. The question of the "human voice" concerns something different in each case. In one, we assess what we received; in the other, what also counts is that this particular person chose to speak. A perfect imitation does not remove this difference. It may, however, make us stop seeing it — and thereby make it easier to sell the appearance of someone's attention.
What Remains of Writing
There is also the author, before anyone reads their work. Linda Flower and John Hayes described writing as a process in which planning, formulating sentences, and evaluating them interweave, and in which a person is capable of changing their own goals along the way. [10] This framing helps to name something invisible in a finished file: part of the thinking happens only during the attempt to write it down in a way that can be defended.
From this perspective, handing a draft to a model may deprive an author of the chance to discover that they don't yet know what they want to say. It may equally create such a chance. Someone else's proposal provokes objection, a well-posed question reveals a gap in the reasoning, and an apt review forces a change of structure. My own example with the second model doesn't permit treating solitary writing as the only worthwhile path. The difference lies elsewhere: between work in which an answer changes my understanding of the matter, and accepting an answer because it is ready and sounds convincing.
For a company buying a report, the difference between these two paths may be invisible: both can yield an equally good document. For the person doing the work, they need not have the same consequences. I'm not claiming that every use of AI weakens the capacity to think; that conclusion doesn't follow from the theory of writing cited above. The point concerns a decision that is easily hidden beneath a speed metric: do we need only the document, or also people who develop the ability to reach its conclusions on their own?
That second goal costs time. Since the tool can also prepare a critique, compare versions, and offer the next question, the efficiency argument may increasingly favour commissioning it to perform the entire sequence of operations. Keeping part of the work on the human side will then require justification extending beyond the quality of any single output. An organisation may wish to develop its employees' competencies, to check someone else's reasoning, to preserve independence of judgement. The word "humanisation" on a list of expectations will not resolve this choice.
When the Author Ceases to Be Necessary
In certain narrow tasks, the difficulty of distinguishing AI creation from human is already a research finding. Let's now adopt a stronger assumption: that in a kind of work that matters to us, a model can for an extended period produce works we judge, on careful reading, to be as good as human ones or better. This extends to wit, to the non-obvious detail, and to a credible point of view. Let's set aside for a moment, then, the consolation that we will always recognise the emptiness beneath the pretty sentences.
What is the author selling then? If a client is buying a result and is indifferent to who prepared it, existing craft proficiency may lose a substantial part of its price. Some commissions will stop going to people. Others will demand less work from them, or work of a different kind. The scale depends on costs, terms of use, and the preferences of recipients — but the mere assurance that humans will shift to "more creative tasks" settles nothing. Creativity is precisely what we placed inside our assumption.
A signature will not restore lost remuneration. Responsibility for the use of a work remains a significant matter, but it doesn't prove that a human can produce something a model cannot. Likewise, an author's biography may increase a work's value for some recipients, without any guarantee that this group will be large enough to sustain existing professions. One can simultaneously value an encounter with a particular human being and commission most utilitarian texts from a machine. We do not know the future proportions of these choices.
Writing may also retain meaning for the writer, even when someone else would do it better. A person writes down their own thought because they want to understand it, to find where they're mistaken, or to leave behind a testimony whose origin is part of its meaning. This value does not depend solely on superiority over a competing producer. It should not, however, be offered as a solution to a professional problem. The pleasure or the need to write and the possibility of making a living from it are different matters.
Let's return to the presentations with the logos covered. At first, their similarity looked like the trace of a shared author. Now one can also see in it the possible trace of a shared commission: explain the world in a way that is easy to accept. A model can do this very well. It can also produce a version that is original and inconvenient. If a company rejects it because it doesn't match the accepted tone, the problem will not be removed by a request for more human character in the next prompt.
An uncovered logo will tell us who approved the document. It will not tell us whether, during its creation, anyone changed their mind, learned something, or had a genuine opportunity to question the commission. In a company that needs people capable of such things, time and space will have to be found for that work even when a finished text can be obtained instantly. This is a cost to be borne consciously, without any promise that the human will prove the better author every time. If all that counts is an efficiently delivered file, it may turn out that the model is entirely sufficient.
Sometimes, however, nobody commissions anything. A person simply wants to share a thought that gives them no peace, though they cannot yet name it well themselves. Something ripens in the recesses of our biological brain until it suddenly takes the shape of an idea from which a text, an image, or a project can emerge. We cannot always reconstruct the route by which it reached us. When we share it with others, someone may take it up, challenge it, or find an application for it we never considered. This, too, is how we learn from one another.
This essay is signed by me, Marcin Górzyński. It began with my need, my questions, and a first sketch of thoughts I wanted to express. AI helped me organise, develop, and test them. I submitted the text for assessment by one model, then another; I commented, selected, rejected, and returned to writing. This is how one can work on any project today — in this case, an article or an essay. What emerges depends also on how much the author brings of their own and how attentively they work with what the tool offers. My AI knows my way of writing and many of my preferences. Very good. But it does not know the thoughts and feelings I haven't disclosed to it. It doesn't even know whether the coffee I drank during this work tasted good to me. I can tell it about that; it will receive only my words. The experience remains mine. I would like a valuable human contribution to meet valuable AI assistance here. Then something better, something stronger, might emerge. Like a good double espresso. Between zero and one hundred percent lies an enormous space that cannot be described by the number of sentences an AI corrected. What we give of ourselves in that space is up to us.
Sources
- Paul J. DiMaggio, Walter W. Powell, "The Iron Cage Revisited: Institutional Isomorphism and Collective Rationality in Organizational Fields," American Sociological Review, 1983. A sociological frame; its application to the circulation of AI content is the author's interpretation.
- Brian Porter, Edouard Machery, "AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably," Scientific Reports, 2024. The 46.6% accuracy figure relates to the first experiment, involving 1,634 participants; quality ratings were studied separately. The result does not settle the question for all genres and audiences.
- J. M. Chein, S. A. Martinez, A. R. Barone, "Human intelligence can safeguard against artificial intelligence: individual differences in the discernment of human from AI texts," Scientific Reports, 2024. The tasks and the earlier model differed from the poetry study; the results should not be compared as though on a single scale.
- Anil R. Doshi, Oliver P. Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances, 2024. Pre-prepared inspirations, without free-form conversation with the model; the study did not test all modes of collaboration with AI.
- Zhivar Sourati et al., "The shrinking landscape of linguistic diversity in the age of large language models," Nature Human Behaviour, 2026. English-language analyses; earlier models were used in the rewriting trials. The observational component relies on an AI-use detector and does not itself establish causation.
- Yun Wan, Yoram M. Kalman, "Diverse AI personas can mitigate the homogenization effect in human-AI collaborative ideation," Computers in Human Behavior: Artificial Humans, 2026. A small experiment; several elements of the procedure were changed, and not all analyses of increased diversity were conclusive.
- Shakked Noy, Whitney Zhang, "Experimental evidence on the productivity effects of generative artificial intelligence," Science, 2023. Results for specified professional tasks, not a measure of overall enterprise productivity.
- Dhruv Agarwal, Mor Naaman, Aditya Vashistha, "AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances," CHI, 2025. Participants from India and the USA wrote in English; the result is not a diagnosis covering all languages and tools.
- Yidan Yin, Nan Jia, Cheryl J. Wakslak, "AI can help people feel heard, but an AI label diminishes this impact," Proceedings of the National Academy of Sciences, 2024. Self-reported sense of being heard in short exchanges was studied, with the effects of content and authorship label measured separately.
- Linda Flower, John R. Hayes, "A Cognitive Process Theory of Writing," College Composition and Communication, 1981. An account of the writing process; its application to collaboration with AI is the author's inference, not a finding of that study.
