Human versus Model: Who Writes Better?

Abstract

Can a generative model genuinely write a better text than a human being, or does its ability to improve existing writing conceal a fundamental dependence upon human authorship? This article transforms that question into a practical literary and visual experiment based not on artificially constructed examples but on a completed and published cultural artefact. The experimental material is issue 45 (Autumn 2026) of the Plovdiv magazine Nula32, devoted to the theme “Time, But Another Time”, whose texts and illustrations were produced entirely by humans before the experiment began. The human original is preserved unchanged and functions as a control version against which a parallel edition produced in collaboration with a large language model can be examined. The principal case study concerns an article on the sociology of literature, the construction of the literary canon and the mechanisms through which certain works acquire recognition and survive across generations, drawing on Pierre Bourdieu and Gisèle Sapiro and incorporating an interview with Alexander Kyosev. The Model is not asked to replace the original subject, research material or human conceptual intuition with an independently invented text. Instead, it is instructed to improve the existing work through clearer composition, more explicit argumentative relations, greater accessibility of theoretical concepts and stronger structural coherence. The experiment consequently distinguishes creation from optimisation and asks whether increasingly radical editorial intervention eventually becomes a form of authorship.

The comparison deliberately refuses to reduce “better” to grammatical correctness, stylistic smoothness or technical sophistication. Conceptual clarity, logical coherence, accessibility, originality, memorability, visual metaphor and the preservation of an individual human voice may not develop in the same direction. The Model can expose latent structures, eliminate redundancy and transform implicit relationships into explicit arguments, yet precisely these abilities generate the experiment’s central paradox: optimisation may improve communicative efficiency while diminishing idiosyncrasy. Repetition may be either redundancy or rhythm; ambiguity may indicate confusion or semantic openness; digression may represent structural weakness or the trace of an individual mind. The visual component produces an analogous problem when technically accomplished generative images are compared with deliberately simpler human illustrations whose value may lie in conceptual economy, humour or metaphor rather than realism. “Human versus Model” therefore does not seek a predetermined winner. It uses the apparent competition to investigate the unstable boundary between author and editor, creation and transformation, defect and style, human intention and machine optimisation. Ultimately, the experiment reveals a third participant whose judgement cannot be eliminated: the Reader. If the human originates the problem, the Model produces the clearer version, and the Reader determines which possesses greater value, the question “Who writes better?” becomes inseparable from a more fundamental one: who decides what “better” means?

Keywords: generative artificial intelligence; large language models; human authorship; literary experiment; editing; human–machine co-authorship; literary value; optimisation; authorial voice; visual communication; generative images; ChatGPT

Introduction: From the Turing Test to the Editor Test — What Does “Better Writing” Mean?

“Machines take me by surprise with great frequency.”

— Alan M. Turing (1950), Computing Machinery and Intelligence

When Alan Turing formulated the famous question “Can machines think?” in 1950, he almost immediately recognised the conceptual trap hidden inside it. Before the question could be answered, one would have to define both “machine” and “thinking”, terms sufficiently unstable to make the argument circular before the experiment had even begun. Turing’s solution was characteristically pragmatic: replace the metaphysical problem with an observable test. Instead of deciding in advance what thinking essentially is, examine whether machine behaviour can become sufficiently similar to human behaviour that an interrogator cannot reliably distinguish between them. More than seven decades later, generative artificial intelligence has displaced this problem into a cultural domain that Turing could only have anticipated indirectly. Whether a machine can produce something recognisable as writing is no longer a particularly provocative question. Large language models routinely generate essays, summaries, advertisements, stories, poems, scripts, translations, reports and innumerable other linguistic forms. The more difficult question begins after this elementary demonstration of competence: can a Model produce a better text than a human being? The apparently innocent adjective “better” transforms a technological comparison into an aesthetic, methodological and philosophical problem. A text can be grammatically impeccable and stylistically lifeless. It can be logically organised yet intellectually predictable. It can become easier to understand while losing ambiguity, rhythm, surprise or individuality. Conversely, a human text may contain repetition, asymmetry, abrupt transitions, syntactic irregularity and partially implicit connections while possessing precisely through these imperfections what readers identify as a distinctive voice. The problem therefore cannot be solved by placing two passages beside one another and counting errors. Writing is not a single competence, and literary quality is not a single measurable variable. Clarity, coherence, originality, accessibility, conceptual precision, memorability and stylistic individuality may reinforce one another, but they may also conflict. The contemporary equivalent of Turing’s problem must consequently ask something more difficult than whether a machine can imitate human writing. It must ask whether the machine can intervene in an already existing human text, identify what prevents that text from realising its own possibilities, improve those elements, and nevertheless recognise what should remain untouched. This article calls that problem the Editor Test.

The experiment begins not with a blank page but with a completed cultural artefact: issue 45 of the Plovdiv magazine Nula32, published in autumn 2026 under the theme “Time, But Another Time”. Its importance as experimental material lies precisely in the fact that its texts and illustrations were created by humans before the comparison with generative AI was conceived. The original therefore cannot retrospectively adjust itself to the Model’s performance. It already exists with its editorial decisions, conceptual strengths, structural weaknesses, visual identity, humour, irregularities and stylistic preferences, and can consequently function as a control version in a limited comparative sense. The principal case examined here is particularly suitable because its own subject mirrors the methodological problem of the experiment. The article considers the sociology of literature, the construction of literary value and the mechanisms through which certain books become recognised, canonised and transmitted to subsequent generations. Pierre Bourdieu and Gisèle Sapiro provide its theoretical coordinates, while an interview with Alexander Kyosev connects those questions to a concrete discussion of literary authority and cultural judgement. The Model receives this already assembled intellectual environment. It does not independently decide to investigate literary value, select Bourdieu and Sapiro, conduct the interview, invent the magazine’s theme or create the cultural context from which the article emerged. Its task is narrower and, for precisely that reason, more revealing: preserve the subject and intellectual material while making the argument clearer, the composition stronger, the theoretical concepts more accessible and the relationships among the existing ideas more explicit. This arrangement deliberately creates an asymmetrical experiment. The human author has already confronted uncertainty, selected the problem and taken the initial creative risk; the Model is permitted to inspect the finished result retrospectively and identify how it might have been organised differently. If its version proves superior according to some criteria, that superiority cannot automatically demonstrate superior authorship. It may instead reveal an extraordinary capacity for optimisation. Yet the distinction becomes unstable as soon as optimisation grows sufficiently extensive. If a Model changes the title, restructures the argument, reorganises examples, rewrites transitions, alters the hierarchy of concepts and replaces substantial portions of the language, at what point does editing become creation? And if the Model makes every relation explicit, every transition smooth and every sentence efficient, what happens to the irregularities through which the original writer remains recognisable? The Editor Test therefore asks two inseparable questions. The first is whether the Model understands what should be changed. The second is more difficult: can it recognise what, despite appearing imperfect, must not be changed?

Part I: The Asymmetrical Duel — Can Human and Machine Ever Compete Fairly?

“The literary field is the site of struggles over the definition of the writer.”

— Gisèle Sapiro (2016), The Sociology of Literature

The decision to use an already published work as the starting point fundamentally changes the meaning of a competition between Human and Model. If both participants were given only a general subject and instructed to produce independent texts, the resulting comparison would immediately become contaminated by uncontrolled differences: each might choose different sources, genres, audiences, examples, argumentative strategies, lengths and even interpretations of the assignment itself. A superficially fair competition beginning from the same prompt could therefore conceal two radically different intellectual tasks. The present experiment deliberately avoids that symmetry. Issue 45 of Nula32 already exists as a completed cultural artefact, produced independently of the subsequent experiment. Its article was not written in order to defeat artificial intelligence, and its illustrations were not designed with the knowledge that generative images would later be placed beside them. This makes the original unusually useful as a control version: its weaknesses cannot be repaired retrospectively once the Model’s alternatives become visible, while its successful decisions cannot be redesigned merely to strengthen the human side of the comparison. Yet this apparent methodological advantage reveals a deeper asymmetry. The Human and the Model are not two writers confronting the same blank page. They enter the process at different moments and inherit different kinds of difficulty. The human author must first recognise that a problem exists, decide that it deserves investigation, choose the intellectual materials through which it can be approached, establish a relationship with a particular audience, conduct the interview, organise the research and transform the magazine’s general theme of time into the specific question of literary survival. The Model enters only after these acts have already taken place. It receives not uncertainty but a solution that can be inspected; not an empty conceptual space but an inhabited one. It can observe the completed architecture from outside and identify where its passages are obscure, where arguments remain implicit, where examples arrive too early or too late, where repetition can be reduced and where an abstract theoretical category could be translated into a more immediate question. The Human therefore possesses the privilege of origin but also bears its risk, while the Model possesses the privilege of retrospection. This difference is not a defect accidentally introduced into the experiment. It is the experiment’s actual subject. The relevant question is not whether artificial intelligence can reproduce the conditions under which a human originally created the article, because those conditions cannot be recreated. The question is what happens when a statistical language model is permitted to enter after the human act of conceptual selection and treat the resulting work simultaneously as finished cultural expression and editable material.

Sapiro’s account of the literary field makes this asymmetry particularly productive because it prevents authorship from being imagined as a private event occurring exclusively between an isolated mind and a blank sheet of paper. Literary works enter networks of editors, publishers, critics, translators, universities, prizes, libraries, booksellers and readers through which cultural legitimacy is produced and contested. Bourdieu’s conception of the field similarly replaces the romantic image of autonomous creation with a space of positions, relations and struggles over different forms of cultural and symbolic capital. Generative AI introduces a peculiar new participant into this already crowded environment. The Model can produce sentences, propose titles, restructure arguments and simulate several editorial functions, but none of these operations automatically answers the question of what kind of participant it has become. If it suggests replacing the descriptive heading “Sociology of Literature” with the more provocative question “Who Decides What Will Remain?”, it has unquestionably altered the communicative power of the article. Yet it has not independently invented the underlying problem. The human original already asks how literary value is established, why some books enter the canon, why others disappear and what social mechanisms intervene between publication and historical survival. The Model recognises that these dispersed concerns can be compressed into a single organising question. Its intervention is therefore neither trivial nor fully originary. Something genuinely changes, but the conceptual material being changed precedes the intervention. The same pattern appears when the Model reorganises Bourdieu, Sapiro, Kafka and Kyosev into a more visible argumentative sequence. The improvement lies not necessarily in adding knowledge but in discovering the latent relations among elements already assembled by the human author. This capacity deserves to be taken seriously because editing has never been intellectually neutral. A strong editor does not merely correct punctuation. Editing can reveal the argument that a writer has imperfectly articulated, eliminate structures that obscure it, identify the sentence around which an entire essay should have been organised and occasionally understand the communicative potential of a manuscript more clearly than its author. The arrival of the Model therefore exposes an old ambiguity rather than creating an entirely new one: authorship has always bordered activities performed by people who were not officially designated as authors. What changes is the scale, speed and linguistic reach with which these interventions can now be performed. A Model capable of restructuring an entire article in seconds forces the cultural field to confront a distinction that conventional publishing could often leave conveniently blurred. Producing language is not identical to originating the reason why that language needed to be produced.

This distinction makes it necessary to dismantle the word “better” into several dimensions rather than allowing it to function as a verdict disguised as a measurement. Conceptual clarity can improve while stylistic individuality declines. Logical continuity can increase while surprise disappears. Accessibility may widen the potential readership while reducing the productive difficulty through which a demanding argument acquires force. Structural economy can remove genuine redundancy but can also remove repetitions that establish rhythm or emphasis. Memorability may arise from perfect formulation, but it can equally arise from an eccentricity that no optimisation system would deliberately preserve. Originality presents an even more difficult problem because the human and machine contributions occupy different causal positions. The Model may formulate a human idea more memorably than its originator, but formulation and conceptual initiative cannot simply be treated as interchangeable. For this reason the experiment evaluates conceptual clarity, logical coherence, accessibility, structural economy, originality, memorability and preservation of authorial individuality as related but potentially conflicting properties. The same principle applies to the visual material. The original Nula32 illustrations use comparatively simple figures and metaphors—books, scales, hierarchies, chess-like relations and visual jokes—to accompany an argument about literary value. A generative system may produce an image with richer texture, more convincing light, greater spatial complexity and apparently superior technical finish. None of those qualities demonstrates that it has produced the better illustration. If the rougher human drawing identifies the intellectual point more precisely, its simplicity may be a strength rather than a deficiency. The controlled comparison therefore produces an unexpected methodological reversal. The more impressive the Model becomes as an optimiser, the less meaningful the original fantasy of an equal duel becomes. The machine is strongest precisely where the Human is most exposed: after the choices have been made, after the risks have been taken and after the imperfect solution has become available for inspection. Its success may consequently demonstrate not that artificial intelligence is a superior human author, but that authorship and editing were never as separable as cultural convention suggested. The Model’s advantage is real, but so is its dependence. It can discover structures the author failed to make visible because the author first created the material in which those structures could be discovered. Human versus Model is therefore asymmetrical by design: the Model receives the extraordinary privilege of seeing what the Human could only discover by writing.

Part II: The Machine as Super-Editor — When Optimisation Begins to Look Like Authorship

“A work of art is a corner of creation seen through a temperament.”

— Émile Zola (1866), Mes Haines

If the first stage of the experiment establishes that Human and Model occupy fundamentally different positions, the second must examine what the Model actually does once it is allowed inside the human text. Here the comparison ceases to be a general discussion about the capacities of artificial intelligence and becomes an anatomy of editorial intervention. The original Nula32 article already contains its essential intellectual materials: the sociology of literature, the problem of literary value, Pierre Bourdieu’s conception of the literary field, Gisèle Sapiro’s analysis of the social conditions of literary production and reception, and Alexander Kyosev’s reflections on the institutions and judgements through which literature acquires cultural authority. The Model therefore does not need to invent a more intelligent subject. Its first significant operation is instead one of recognition: among the many questions contained in the human text, which one can become the gravitational centre around which everything else is reorganised? The apparently modest transformation of the title “Sociology of Literature” into “Who Decides What Will Remain?” illustrates the process. The first title names an academic discipline; the second exposes a problem. This alteration does not add a new theory, source or fact, yet it changes the reader’s route through the entire argument. Instead of being invited to learn what the sociology of literature is and only later discover why it matters, the reader encounters an intuitively comprehensible question and then discovers that sociology is necessary because the obvious answer—“time will decide”—is insufficient. The machine-assisted version can consequently make explicit an argument that existed only partially articulated in the original: time does not read books. Time does not preserve manuscripts, commission translations, publish editions, award prizes, construct university syllabuses, maintain libraries or persuade one generation that a writer deserves to be read by the next. People and institutions perform these actions. What retrospectively appears to be the impartial “judgement of time” is therefore produced through countless acts of selection, preservation, interpretation and transmission. The Model’s achievement is revealing precisely because it is difficult to classify. It has not discovered a new sociological law. It has identified a latent argumentative structure within human material and made that structure visible. The machine appears strongest here not as an autonomous originator but as a reader capable of returning the human text to its author in a form that exposes what the author was already trying to say.

The example of Franz Kafka makes this editorial operation especially concrete. Kafka’s position within the European literary canon can appear inevitable only when viewed backwards from a cultural world in which his name has already become an adjective, his works have been translated into numerous languages, and his place within academic and public culture has been institutionalised. Yet the survival of the work depended upon particular acts of mediation, most dramatically Max Brod’s decision not to destroy Kafka’s unpublished manuscripts. The example converts the abstract sociology of cultural survival into a simple counterfactual proposition: if the manuscript disappears, “time” has nothing to judge. From that point the Model can organise material that the human original presents more diffusely into a chain of cultural transmission. Someone preserves the manuscript; another person publishes it; another translates it; critics interpret it; institutions legitimate it; libraries conserve it; teachers transmit it; readers recommend it. Sapiro’s categories of production, circulation and reception cease to function merely as theoretical vocabulary and become observable stages through which literary survival is produced. Bourdieu’s literary field similarly becomes easier to grasp when represented not as an abstract diagram of positions but as a network of actors whose judgements distribute different forms of cultural recognition. The Model performs a comparable operation when it separates four categories easily confused in ordinary discussion: the good book, the popular book, the institutionally recognised book and the book that survives. These categories can overlap, but they need not coincide. Commercial success does not guarantee canonical status; academic prestige does not require mass readership; an initially neglected work may later become central; a bestseller may disappear rapidly from cultural memory. Once again, the machine has not invented the distinction from nothing. It has converted an implicit structure into an explicit analytical sequence. This suggests that one of the most powerful functions of generative editing may be neither information retrieval nor sentence correction but conceptual externalisation. The Model can detect relationships dispersed across a manuscript and return them as an architecture. In doing so, however, it begins to exceed the conventional image of an editor as someone who merely cleans the surface of a completed work. If it decides which question should organise the article, which examples should carry the theoretical burden, which distinctions require explicit formulation and in what sequence the argument should unfold, it is intervening in the organisation of meaning itself. The boundary between editing and authorship consequently becomes difficult to locate not because the distinction has disappeared, but because both activities increasingly occupy the same textual territory.

The difficulty becomes still greater when the experiment turns from what optimisation gains to what it removes. The Model is remarkably effective at detecting repetitions, weak transitions, delayed theses, unexplained theoretical terminology and structural disproportion. It can move a proposition forward, place an example immediately after the claim it illustrates, compress several similar sentences into one, convert an implicit relationship into an explicit causal connection and replace specialised terminology with an accessible explanation. For journalistic and public-facing writing these interventions can be extraordinarily useful. Yet the logic that produces them contains a literary danger. Not every repetition is redundant; repetition can create rhythm, insistence or rhetorical pressure. Not every digression is evidence of failed organisation; a deviation can reveal the associative movement of an individual mind. Not every ambiguity represents conceptual weakness; ambiguity may deliberately preserve several meanings that a more efficient formulation would prematurely collapse. Even syntactic awkwardness can sometimes function as an identifying feature rather than an error. Zola’s “temperament” becomes crucial at precisely this point. If a work is creation refracted through a particular temperament, then editing cannot be understood solely as the elimination of deviation from an ideal communicative norm. Some deviations are the temperament. A system optimised towards clarity, coherence and probable linguistic appropriateness therefore confronts a problem more difficult than grammatical correction: it must distinguish defect from signature. The visual experiment produces the same dilemma in immediately visible form. A generated illustration may contain richer textures, more sophisticated lighting, greater realism and a more polished composition than the original human drawing, yet still be inferior as an illustration if it misunderstands the intellectual metaphor that connects image and text. Technical finish cannot substitute for conceptual necessity. The same is true of prose. A beautiful sentence is not automatically the necessary sentence, and a flawlessly organised paragraph may not be the most memorable one. The Model’s greatest editorial strength—the capacity to reduce uncertainty and rapidly impose structure—can therefore become its greatest creative weakness when optimisation begins to standardise the very irregularities through which a particular voice becomes recognisable. At that threshold the question “Can the Model improve the text?” is no longer difficult enough. The experiment must ask something more disturbing: how much human imperfection can be removed before the human author begins to disappear with it?

Part III: The Right to Be Imperfect — When Better Writing Becomes Less Human

“Style is a way of saying who you are without having to speak.”

— Rachel Cusk (2014), Outline

The experiment reaches its most difficult stage precisely when the Model succeeds. If the machine-assisted version becomes clearer, more coherent, more economical and more accessible than the human original, it is tempting to treat the result as evidence that the Model has written the better text. Yet such a conclusion immediately encounters the unresolved problem of what exactly has been compared. The Model did not independently decide that literary survival was worth investigating; it did not select the sociology of literature as the conceptual framework; it did not choose Bourdieu and Sapiro; it did not conduct the interview with Alexander Kyosev; and it did not originate the editorial environment of Nula32 from which the article emerged. Those acts precede its intervention. What the Model receives is already an organised field of human intention, research and cultural judgement. When it replaces “Sociology of Literature” with “Who Decides What Will Remain?”, it does not create a new intellectual object so much as discover a more powerful formulation of an existing one. When it separates quality, popularity, institutional recognition and historical survival into distinct categories, it makes explicit a conceptual difference already latent in the original material. Nevertheless, describing these interventions merely as correction becomes increasingly implausible. An editor who changes the organising question, rearranges the conceptual hierarchy, determines the sequence of examples, reconstructs transitions and rewrites substantial portions of the prose participates in the production of meaning, even if the initial reason for producing that meaning belongs elsewhere. Generative AI therefore destabilises the conventional boundary between author and editor not by proving that the distinction is meaningless, but by demonstrating how many different activities had previously been compressed into the single word “writing”. Conceiving the problem, selecting the material, producing sentences, arranging arguments, finding metaphors, revising formulations and judging the final result are separable operations. Traditionally they could all occur inside one human consciousness and therefore appear to constitute a unified act of authorship. The Model allows some of these operations to migrate outside that consciousness. The resulting text may consequently possess several causal origins at once: a human reason for existing, human research material, machine-generated formulations, machine-assisted structure and finally human acceptance or rejection. “Who wrote it?” becomes difficult to answer because writing itself has been decomposed.

The more consequential problem, however, concerns not the ownership of successful improvements but the status of the imperfections they remove. Editing conventionally assumes that a text contains features that interfere with the text it ought to become. The editor therefore searches for redundancy, inconsistency, obscurity, weak organisation, inappropriate diction and unnecessary complexity. A generative Model can perform many of these operations with remarkable speed because large-scale statistical regularities allow it to recognise probable patterns of effective exposition. But literary individuality often emerges precisely where an individual departs from probable patterns. A repeated phrase may be unnecessary information and simultaneously necessary rhythm. A sentence that appears excessively long may reproduce the pressure of a thought refusing premature closure. An abrupt transition may indicate poor composition, or it may create a productive collision between ideas. An unexplained image may confuse one reader while remaining unforgettable to another. Ambiguity can be an error when the writer does not know what is being said, but it can also be the mechanism through which a text refuses to say only one thing. The central editorial problem therefore becomes a problem of classification: which irregularities are defects and which are signatures? Human editors confront the same difficulty, but generative optimisation intensifies it because the Model can revise at a scale and speed that encourage the fantasy that every detectable irregularity ought to be resolved. The danger is not necessarily that AI writes badly. The more interesting danger is that it may write too recognisably well. If clarity, balance, transition, explanatory completeness and stylistic smoothness are continuously rewarded, texts can converge towards an increasingly familiar morphology of competent prose. Nothing is obviously wrong, but increasingly little resists expectation. The paradox of improvement can therefore be formulated precisely: a text may become better according to numerous formal criteria while becoming less valuable as evidence of a particular human mind. This does not make imperfection intrinsically virtuous. Confusion remains confusion; cliché does not become originality merely because a person produced it; incoherence should not be protected under the romantic name of individuality. The task is subtler. Human authorship requires the right to certain non-optimal choices because optimisation cannot know in advance which deviations will later be recognised as style. A literature from which every statistically improbable roughness has been removed might be extraordinarily readable and culturally sterile at the same time.

The visual component of the experiment makes this paradox almost impossible to ignore. A generative image can immediately surpass a simple human illustration in texture, lighting, depth, anatomical plausibility and technical finish. Yet illustration is not a competition in rendering capacity. Its purpose is to establish a meaningful relation with the text. A crude drawing containing the correct metaphor can therefore outperform a visually spectacular image whose relationship to the argument is generic. The same principle returns writing to the problem of authorship. Technical competence is not identical to conceptual necessity. The sentence that sounds best in isolation may not be the sentence the work requires, just as the most beautiful image may not be the image that thinks most precisely beside the article. This is where the experiment finally exceeds its original opposition between Human and Model. A third participant has been present from the beginning: the Reader. Neither the human author nor the generative system can independently convert its own criteria into a final judgement of literary value. The Model may identify its version as clearer because its argument has become more explicit; the author may prefer the original because it preserves a particular rhythm or uncertainty; neither judgement settles whether the intervention produced a better cultural object. The Reader encounters the result and decides whether the removed irregularity was noise or voice, whether the clarified ambiguity was confusion or richness, whether the improved structure reveals the argument or makes it excessively predictable. This does not make literary value arbitrary. It reveals instead that value is relational and historically unstable, precisely as the sociology of literature underlying the original Nula32 article suggests. Institutions, critics, editors, readers and subsequent generations continually renegotiate which qualities matter. Generative AI now enters that field not merely as another producer of cultural objects but as an instrument capable of modifying the objects before judgement begins. The initial duel consequently becomes a triangle: Human, Model, Reader. And once the Reader enters, the possibility of a simple technological victory disappears. The Model can produce the clearer text. The Human can originate the problem and preserve the more distinctive voice. The Reader may still prefer either version—or discover that the most successful text is neither, but a hybrid in which machine optimisation has been deliberately stopped before it eliminates the human irregularity worth preserving. The deepest question raised by the experiment is therefore not whether artificial intelligence can become sufficiently human to write literature. It is whether future writers will remain sufficiently willing to resist optimisation to preserve those forms of imperfection from which recognisable human style can emerge.

Conclusion: Who Actually Won? — The Problem of the Jury

“A work is never completed except by some accident such as weariness, satisfaction, the need to deliver, or death.”

— Paul Valéry (1933), “Au sujet du Cimetière marin”

The experiment Human versus Model began with a deliberately simple question: who writes better? By its conclusion, the difficulty lies less in producing the competing texts than in preserving any stable meaning for the words “writes” and “better”. The human version of the Nula32 article supplied far more than raw sentences awaiting technical improvement. It supplied the reason for the text to exist: the initial problem, the intellectual curiosity, the theoretical sources, the interview with Alexander Kyosev, the editorial circumstances, the relationship with the magazine’s theme and the decision to investigate literary survival through the sociology of literature. The Model entered this already inhabited conceptual environment and performed a different kind of intellectual labour. It read retrospectively. It identified a stronger organising question, exposed relationships that remained partially implicit, distinguished categories that the original allowed to overlap, rearranged argumentative sequences and converted specialised theoretical material into a more immediately accessible public form. Judged exclusively by conceptual transparency, structural coherence and communicative economy, such interventions can constitute substantial improvements. Yet their success does not establish that the Model independently authored the intellectual object it improved. The machine does not optimise a void. It optimises a prior arrangement of human intention. At the same time, calling its contribution mere proofreading would be equally misleading. Once an intervention changes the title, conceptual hierarchy, architecture, transitions and substantial portions of the language, the traditional distinction between author and editor begins to lose its apparent simplicity. The experiment therefore suggests that generative AI is valuable precisely because it makes visible something that literary culture could previously conceal: authorship has always consisted of separable operations. Conceiving, researching, selecting, composing, restructuring, revising and judging do not necessarily belong to the same agent. Generative systems allow these operations to be distributed between biological and computational participants with unprecedented speed and fluidity. The resulting cultural object can consequently remain human in origin while becoming partly machine-mediated in expression. The important question is no longer whether such a text is simply “human” or “AI-generated”, but which participant performed which transformation and which decisions remained indispensable to the final work.

The more unexpected result concerns the limits of improvement itself. The experiment provides no universal measure capable of converting literary quality into a single quantity. A text can become clearer while becoming more predictable, more coherent while becoming less idiosyncratic, more elegant while losing an awkward but memorable formulation. The same applies to images: technical sophistication cannot guarantee metaphorical precision, just as visual realism cannot replace the intellectual decision that makes an illustration belong beside a particular text. Optimisation and creative value are therefore not synonyms. This distinction becomes especially important because literature does not always succeed by transmitting information with maximum efficiency. Some texts remain culturally productive precisely because they cannot be exhausted by an optimal paraphrase. They preserve ambiguity, contradiction, rhythm, silence, excess and interpretative resistance. A machine trained to identify probable improvements may be extraordinarily useful when a writer has failed to express an intended relationship, but the same capacity becomes dangerous when it assumes that every irregularity is a failure awaiting correction. Human authorship therefore cannot be defended merely by celebrating imperfection. Bad writing remains possible, and human origin does not magically transform confusion into style. The more defensible principle is that creative work requires the possibility of non-optimal decisions. A writer must be allowed to repeat where compression would be more efficient, to withhold where explanation would be clearer, to choose an eccentric metaphor where a conventional one would communicate faster, and occasionally to preserve a sentence that an excellent editor—human or artificial—would remove. Some of these decisions will fail. Others may become the very features through which the work acquires a recognisable identity. The paradox is that no optimisation system can determine this distinction with certainty beforehand because literary history itself repeatedly changes the criteria through which eccentricity becomes innovation and difficulty becomes value. The right to be imperfect is therefore not an argument against editing. It is a defence of the residual uncertainty without which editing could become standardisation.

The experiment consequently ends with a participant who initially seemed external to the contest but proves necessary to it: the Reader. Human and Model can produce different versions, but neither possesses an independent authority to declare what “better” ultimately means. The human author may value intention, originality and the preservation of voice; the Model may produce a version that demonstrably increases clarity, coherence and accessibility; an editor may favour economy; a critic may value formal innovation; one reader may prefer the polished version while another remembers precisely the strange sentence that optimisation removed. This does not reduce literary judgement to arbitrary preference. It returns the experiment to the sociological problem from which it began. Cultural value emerges through relations among works, readers, institutions, critics, publishers and historical circumstances. Generative AI now enters this field as another powerful mediator, but it does not abolish the field itself. The apparent duel Human versus Model therefore becomes a triangle: Human → Model → Reader. More accurately, it becomes an expanding network in which authorship, editing, interpretation and evaluation can no longer be assigned automatically to a single participant. The future of writing may consequently belong neither to the untouched human manuscript nor to the autonomously generated machine text, but to an unstable collaboration in which the decisive creative act sometimes consists in knowing when to accept the Model’s intervention and when to refuse it. The best human–machine text may not be the one in which artificial intelligence has optimised everything it can detect. It may be the one in which optimisation is deliberately stopped at the point where further improvement would begin to erase the person whose text was supposed to be improved. Human versus Model therefore cannot announce a universal winner without contradicting its own findings. If the Model produces the clearer version, the Human originates the problem, and the Reader determines which version possesses greater value, the original question has changed. The issue is no longer simply who writes better. It is who creates, who transforms, who judges—and, ultimately, who is allowed to decide what “better” means.

References

Ayolov, P. (2026) The Conspiracy of Speech. Vol. I of The Planned Obsolescence of Language: Miscommunication Trilogy.

Ayolov, P. (2026) The Entropy of Communication. Vol. II of The Planned Obsolescence of Language: Miscommunication Trilogy.

Ayolov, P. (2026) Negentropic Culture: Against the Entropy of Words. Vol. II, Part 2 of The Planned Obsolescence of Language: Miscommunication Trilogy.

Ayolov, P. (2026) The Tower of Babble. Vol. III of The Planned Obsolescence of Language: Miscommunication Trilogy.

Ayolov, P. (2026) ‘The Death of the Reader 2.0: From Roland Barthes’ Scriptor to the AI Prompter’.

Ayolov, P. (2026) ‘Language as a Tissue of Citations: How a Self-Quoting Mechanism Conquered the Human Mind’.

Ayolov, P. (2026) ‘The Neo-Awakening of Information: The Illusion of Self, LLMs, and the Preservation of the Immortal Pattern’.

Ayolov, P. (2026) ‘Automating the Wilde–Bayard Paradox: LLMs, Predictable Equilibrium and the Death of Deep Listening’.

Ayolov, P. (2026) ‘From Snubbing to Znubbing: The Amputation of Attention’.

Ayolov, P. (2026) ‘Human versus Model: Who Writes Better? A Literary Experiment in Human Authorship, Generative Artificial Intelligence, and the Boundary between Editing and Creation’.

Barthes, R. (1977 [1967]) ‘The Death of the Author’, in Heath, S. (trans.) Image–Music–Text. New York: Hill and Wang, pp. 142–148.

Boden, M. A. (2004) The Creative Mind: Myths and Mechanisms. 2nd edn. London: Routledge.

Bourdieu, P. (1993) The Field of Cultural Production: Essays on Art and Literature. Edited and introduced by R. Johnson. Cambridge: Polity Press.

Bourdieu, P. (1996) The Rules of Art: Genesis and Structure of the Literary Field. Translated by S. Emanuel. Stanford, CA: Stanford University Press.

Calvino, I. (1999 [1991]) Why Read the Classics? Translated by M. McLaughlin. London: Jonathan Cape.

Cusk, R. (2014) Outline. London: Faber & Faber.

Eco, U. (1979) The Role of the Reader: Explorations in the Semiotics of Texts. Bloomington: Indiana University Press.

Eco, U. (1989) The Open Work. Translated by A. Cancogni. Cambridge, MA: Harvard University Press.

Foucault, M. (1977 [1969]) ‘What Is an Author?’, in Bouchard, D. F. (ed.) Language, Counter-Memory, Practice: Selected Essays and Interviews. Translated by D. F. Bouchard and S. Simon. Ithaca, NY: Cornell University Press, pp. 113–138.

Good, I. J. (1965) ‘Speculations Concerning the First Ultraintelligent Machine’, in Alt, F. L. and Rubinoff, M. (eds) Advances in Computers. Vol. 6. New York: Academic Press, pp. 31–88.

Iser, W. (1978) The Act of Reading: A Theory of Aesthetic Response. Baltimore: Johns Hopkins University Press.

Manovich, L. (2001) The Language of New Media. Cambridge, MA: MIT Press.

McLuhan, M. (1964) Understanding Media: The Extensions of Man. New York: McGraw-Hill.

Richards, I. A. (1924) Principles of Literary Criticism. London: Kegan Paul, Trench, Trubner & Co.

Sapiro, G. (2016) The Sociology of Literature. Translated by M. Ryan. Stanford, CA: Stanford University Press.

Turing, A. M. (1950) ‘Computing Machinery and Intelligence’, Mind, 59(236), pp. 433–460.

Zola, É. (1866) Mes Haines: Causeries littéraires et artistiques. Paris: Achille Faure.

Nula32 (2026) Vremeto, ama drug pat [Time, But Another Time], No. 45, Autumn 2026. Plovdiv: Nula32.

Kyosev, A. (2026) ‘Interview’, in Vremeto, ama drug pat [Time, But Another Time], Nula32, No. 45, Autumn 2026. Plovdiv: Nula32.