EDITORIAL · Editorial

Perfectly Boring: Why AI’s ‘Best’ Writing is Fundamentally Unreadable

AI can produce competent, even affecting prose. Its default polish is still a problem: when every sentence sounds carefully reasonable, nothing in the paragraph feels chosen at a cost.

By The Report Editorial Board · October 29, 2026

This is an editorial judgment informed by five linked research sources. The studies support bounded claims about decoding, in-context task specification, instruction tuning, and creative-story diversity; they do not prove that every AI passage is predictable or that surprise is the only literary virtue.

A sentence that survives every substitution has probably escaped the circumstances that would make it matter.

The passage arrives aggressively clean. It acknowledges complexity, promises analytical depth, and glides toward a conclusion with the inoffensive manners of a corporate lobby. No sentence is technically wrong. No paragraph trips over its own syntax. Still, a practiced reader feels the urge to interrupt. The prose has given the reader no occasion to lean forward, argue back, or wonder why this sentence had to be written in this form.

That reaction is not a complaint that writing must be ragged. Clarity is a virtue, and many assignments call for tact, speed, or a plain answer. A customer-service reply should not arrive with a novelist’s weather. A grant summary has obligations that a lyric essay does not. The trouble begins when the manner suitable to routine explanation becomes the assumed manner for every subject, including grief, conflict, history, and art. Then fluency starts to resemble thought while quietly substituting for it.

By unreadable, I do not mean illegible, false, or incapable of affecting anyone. A generated passage can soothe, amuse, persuade, or provide a useful starting point. I mean unreadable in the strict reading sense: prose that asks for attention without returning much of it. It moves smoothly from premise to conclusion, but its smoothness has no pressure behind it. A reader can recognize every gesture before the paragraph has made it.

The etiquette of the default

Default AI prose often begins by widening the frame. It grants the obvious counterpoint, arranges clauses with admirable courtesy, and ends on a conclusion no reasonable person could reject. This is not a mysterious flaw. It is the rhetoric of low-risk completion. The wording keeps doors open because open doors make fewer immediate demands on the person who requested the text. In institutional language, that discretion can be useful. In criticism or narrative, it can turn a subject into upholstered furniture.

Consider this invented illustration, not a collected AI sample: “As cities reconsider public space, the shift offers valuable opportunities while raising important questions that call for a balanced response.” The sentence is grammatical, calm, and nearly empty. It does not name which space is contested, who gains access, who loses a bench or a bus stop, or what decision must be made. Replace “cities” with schools, publishers, or hospitals and the sentence still functions. That portability is the problem. A sentence that survives every substitution has probably escaped the circumstances that would make it matter.

Concrete diction is not the same thing as decorative detail. “A family waited forty minutes for the only evening bus” may belong in an argument about transit because it establishes a human timetable and a material consequence. “The urban mobility ecosystem faces evolving challenges” does not become better because it has more syllables. One phrase identifies a condition someone could dispute, investigate, or remedy. The other protects the writer from having to decide what the condition is.

Human beings produce this evasive language too. Press offices, committee reports, and anxious executives have long rewarded sentences that can survive a room full of objections. The novelty is scale and convenience. A system that can produce polished generalities on demand makes it easier to stop before anyone asks the harder question: what, in this paragraph, could not have been said by anyone else about anything else?

Prediction is not an average word

The common explanation of this blandness is wrong. A language model does not simply select an average next word, as though “average” were a token available on a shelf. Given a context, it assigns probabilities across possible continuations. A decoding method then helps determine what comes next: it may favor high-probability tokens, sample among a restricted set, or apply other controls. The same model can produce notably different text under different decoding choices. Holtzman and colleagues’ study of neural text degeneration [1] showed that likelihood-driven decoding could produce bland or repetitive output and that decoding strategy materially affected diversity, fluency, and coherence.

That technical distinction matters because it prevents a lazy fatalism. Flat prose is not an unavoidable mathematical average, and a surprising phrase is not proof of authorship. Sampling can make a continuation less predictable; a writer can ask for a changed register; a prompt can establish a task, examples, and constraints. In the GPT-3 experiments described by Brown and colleagues [2], instructions and demonstrations supplied in text conditioned performance without retraining the model. Context changes the available surface.

Instruction tuning changes the surface further. Ouyang and colleagues [3] trained InstructGPT with human demonstrations and ranked responses, and on their evaluated prompt distribution human labelers preferred a smaller instruction-tuned model to the larger base GPT-3. That result is worth taking seriously: alignment to human preferences can improve usefulness. It does not mean a preference-trained voice has inherited the position from which a person makes claims, retracts them, or bears the cost of being wrong.

The concern, then, is less about a machine’s supposed taste than about what a default setting rewards. Probability and preference tend to favor continuations that look acceptable across many plausible contexts. A writer sometimes needs the opposite impulse. She may need the stubborn noun, the unflattering sequence of facts, the sentence whose cadence tightens because a concession has run out. A system can supply candidates for those moves. It cannot make the risk socially real on its own.

When polish becomes a house style

Polish is often mistaken for rhythm. It is not. Rhythm is the management of expectation through sentence length, stress, syntax, and pause. An eight-word sentence after a long one can expose a fact. A subordinate clause can delay judgment until evidence arrives. Repetition can be comic, accusatory, or mournful when it gathers force from a particular scene. Default polished prose tends instead toward even pacing: qualifications arrive on schedule, transitions explain themselves, and each paragraph lands with a tidy abstraction. The effect is less a voice than a house style.

That house style is especially damaging when it handles difficult material. A paragraph about eviction that says “housing insecurity creates complex challenges” has adopted a neutral altitude before it has named the landlord, the court date, the rent increase, or the child changing schools. A paragraph about a public figure’s failure that calls for “a nuanced conversation” may be correctly cautious, yet it can also use caution to postpone judgment forever. Precision in prose begins with reference. Who did what, to whom, under what constraint, and with what result?

Strong prose does not demand that every sentence be flashy. It drops a vague intensifier because the word blurs the claim. It asks whether a metaphor discovers an angle or merely decorates a familiar thought. It notices when three balanced clauses have marched the paragraph into sleep. Often, subtraction makes room for the phrase with a job to do and preserves the awkward fact that would otherwise be sanded into a statement of principle.

Default polish has another cost: it trains writers to confuse finish with decision. A paragraph may have transitions, varied verbs, and a concluding sentence, yet remain undecided about its subject. It may use “however” to simulate opposition without placing two claims in conflict. It may call an issue “complex” as a substitute for showing which values collide. Such prose looks resolved from a distance. At close range, its central noun has not been earned.

Diversity studies are not a taste tribunal

There is empirical reason to take sameness seriously, although the evidence should not be inflated into a universal rule of literary value. In a 2025 study of 100 story continuations, Xu and colleagues [4] found recurring, idiosyncratic plot elements in outputs from GPT-4 and LLaMA-3 that did not recur in the original human stories in their set. Their automatic Sui Generis score correlated moderately with human surprise judgments. The study concerns particular prompts, models, and story generation. It does not establish that every generated paragraph is predictable, nor that surprise is the only virtue a story can possess.

A second comparison clarifies the limits. In a five-sentence cue-word story task, Ismayilzada, Stevenson, and van der Plas [5] reported that LLM stories were stylistically complex but lower than average human writers on automated measures of novelty, surprise, and diversity. Expert ratings generally tracked those measures, while non-experts and LLM evaluators rated the LLM stories more creatively. This disagreement matters. Readers bring different expectations, and a quiet story may be better than an ingenious one. No score can settle the question of whether a sentence has earned its place.

Still, the findings describe a practical writing risk. If a tool returns familiar reversals, reconciliations, and images across many requests, the writer must look harder for the choice that belongs to this work. Novelty for its own sake will not solve the problem. Randomness is not revelation. Test a seemingly apt line against the alternatives it has displaced. Does the detail alter our understanding of the person or problem? Does the ending follow from what came before, or has it been selected because it resembles an ending?

Imitation and authorship are different claims

A model can imitate features associated with a style: a clipped sentence, a favorite adjective, a pattern of dialogue, a familiar turn from observation to judgment. Those features are real and can be useful for an exercise, a parody, or a constrained commission. But resemblance is not yet authorship. Authorship involves responsibility for selection. Someone has to decide that this comparison is fair, that this accusation is warranted, that this omission is defensible, and that this risk is worth taking before an audience that may answer back.

The distinction is clearest when prose goes wrong. A writer can be asked why she chose a euphemism, why she framed a neighborhood through danger, why she omitted the person most affected by a policy. She may defend the choice, apologize, or accept the damage to her reputation. A publisher, working with the writer, can make those judgments too. Fluent text does not remove those obligations. It merely makes it easier to conceal that nobody has fulfilled them.

This is why the relevant question is not whether a generated passage can resemble literary writing. It can, and sometimes it can do so with considerable skill. The question is whether a responsible person has made the passage answerable to a situation. Lived stakes need not mean confession or autobiography. A historian has stakes in evidence; a critic has stakes in fairness; a reporter has stakes in accuracy; a novelist has stakes in what a made-up life permits a reader to feel and infer. Each discipline binds style to a consequence outside the sentence.

Writing under consequence

The remedy is not a purity ritual. A writer may use a model to test an outline, surface an overlooked objection, compare possible structures, or generate a clumsy beginning that reveals the shape of a problem. Those uses can save time. The error comes when a finished-looking answer is treated as evidence that the underlying choices have been made.

Before accepting a polished paragraph, a writer can ask more demanding questions. What is the claim that must remain after the cushioning language is removed? Which noun identifies the actual object of dispute? Where does a sentence rely on “important,” “complex,” or “meaningful” because the writer has not supplied evidence? Who could reasonably object, and has the objection changed the argument rather than merely been acknowledged? These are not tricks for producing a recognizable voice. They are methods for making prose answerable.

The strongest choice may be a cut. Remove the sentence that announces the coming argument and begin where the argument begins. Replace the general scene with one verifiable condition. Let a conclusion remain limited when the evidence is limited. Keep a strange sentence when its strangeness comes from accurate attention, not from a bid to appear original. The aim is neither maximal smoothness nor ornamental difficulty. It is prose in which the reader can feel a mind selecting, refusing, and accepting the consequences of what it says.

Opinion informed by linked research. Interpretations and conclusions are editorial judgments.