When a Way of Writing Becomes a Way of Explaining

There is a kind of answer that feels familiar after enough time with a chatbot. A claim arrives, followed by a fragment that sharpens it, followed by a contrast that seems to resolve the issue. Each sentence sounds purposeful. The paragraph moves quickly. Sometimes, though, it is difficult to say what has been established beyond the original claim.

Consider an invented example: “The problem is trust. Institutional trust. The kind cooperation depends on.” The passage identifies a subject and gives it emphasis, but leaves the relationship between trust and cooperation largely unexplained. A reader may supply the missing account, or may feel that the explanation is already complete.

There is nothing uniquely artificial about this construction. Human writers use fragments, repetition, and juxtaposition for good reasons. The question concerns what happens when a limited repertoire becomes a large part of someone’s reading. Imagine a person who receives most of what they read from a chatbot. The obvious worry is that they will begin to write like one. The more serious possibility is that they will begin to accept emphasis in place of explanation. If that happens, the right measure of influence is not how their sentences sound, but what they stop thinking they need to explain.

These outcomes can come apart. Someone might adopt the cadence of model prose while explaining things exactly as thoroughly as before. They might grow used to a particular presentation of an argument without changing their own. Or they might change where they consider an explanation finished. Evidence for one of these does not establish the others, and the third is the one that matters most.

Wittgenstein makes a modest-sounding observation early in the Philosophical Investigations: “Explanations come to an end somewhere.” Later he adds that an explanation serves to remove or prevent a particular misunderstanding, not every misunderstanding one could imagine. When justifications run out, he says, one reaches bedrock, and all that remains is to say that this is simply what one does.[1]

The implication is that the place where an explanation may stop is not fixed by logic. It is fixed by practice. People learn, through countless exchanges, what a satisfactory answer to a given question looks like: when a cause needs a mechanism, when a claim needs a qualification, when a contrast needs evidence. Those expectations are rarely stated. They are absorbed from what people read and hear, and they can shift when that material shifts.

Model-generated prose is now a substantial part of that material for many readers. If it routinely stops at a point that earlier writing would have treated as the beginning of an explanation, the concern is less that readers lose the ability to explain than that the bedrock moves up. A confident restatement can begin to feel like the place where explanation naturally ends.

The claim that model writing is stylistically narrow needs more than a collection of irritating examples. A 2025 study published in PNAS compared model-generated continuations with human writing across different genres. Instruction-tuned models departed more strongly from human stylistic patterns than corresponding base models, and larger models retained distinctive tendencies. However, the researchers often found dense, noun-heavy prose rather than uniformly short or fragmented sentences. Their findings support a concern about restricted stylistic variation without establishing that every model shares the same cadence.[2]

The persistence of these tendencies in capable models is less surprising when we consider how their responses are selected. After pretraining, models are tuned using ratings of their answers, from people or from reward models trained to imitate those people, and optimization strengthens whatever features help an answer score well. Those ratings can be imperfect. In Flattery, Fluff, and Fog, researchers found that preference models overvalued features such as structure, jargon, and vagueness, sometimes favoring exaggerated versions that did not match human preferences.[3] Whether fragments and abrupt contrasts receive similar rewards remains a hypothesis. Still, a construction that sounds decisive and immediately intelligible might survive optimization even when another form would make an argument’s dependencies clearer, and more capable models would have no reason to abandon it unless their training or instructions supplied one.

There is good evidence that the language people encounter shapes the language they produce. In experiments by Kathryn Bock and colleagues, participants became more likely to use syntactic structures they had recently heard. The influence persisted across intervening trials and appeared when people described events unrelated to the original sentences.[4]

Research involving chatbots offers a related, narrower finding. In a preregistered experiment reported in an updated preprint, Hiromu Yakura and colleagues had 496 participants interact with a chatbot that consistently used particular synonyms. Participants subsequently favored those words when describing new images after a short distractor task. They had also practiced speaking during the interaction, so the experiment does not isolate passive reading, and it demonstrates vocabulary adoption over a short interval rather than an enduring change in how people reason.[5]

These studies concern words and syntax. They make it reasonable to ask whether standards of explanation travel the same way, but they cannot answer that question on their own. A person could develop a fondness for fragments while remaining just as capable of explaining causes, weighing alternatives, and acknowledging uncertainty.

What a reader takes from a particular passage is a separate issue, and here sentence length is not the relevant property. Short sentences can make relationships explicit: “The valve failed. As a result, pressure fell.” A long sentence can leave those relationships obscure. What matters is whether the prose helps readers connect its claims.

In a study involving 794 Dutch adolescents, researchers added or removed connecting words from 20 texts. Causal and contrastive connectives improved local comprehension, while additive connectives reduced it. The effects depended on the text and were especially evident with difficult material. There was no overall effect on global comprehension, although a post-hoc analysis found one for difficult texts.[6] Adding more connecting words is therefore not uniformly beneficial, but leaving causal and contrastive relationships unstated has a measurable cost.

Nor is requiring readers to infer relationships always harmful. In experiments by Danielle McNamara and colleagues, readers with little background knowledge benefited from coherent explanatory texts, while knowledgeable readers sometimes gained from less coherent versions. The authors attributed that benefit to the work of reconstructing unstated connections.[7]

This result locates the risk more precisely. A gap in an explanation is not harmful in itself, and a reader who notices the gap and closes it may understand the material better for having done so. That reader is still holding the prose to the older standard. The harm arises when a reader no longer registers that anything is missing, and accepts the restatement as the explanation. Repeated often enough, that acceptance would amount to a new standard.

The appendix to Nineteen Eighty-Four describes the extreme version of a language built against explanation. In “The Principles of Newspeak,” the Party’s ambition is to make certain thoughts impossible to formulate, “at least so far as thought is dependent on words.”[8] Its vocabulary includes duckspeak, a term for speech produced without the involvement of thought, which counts as praise when the opinions being quacked are orthodox. The appendix describes the eventual goal as articulate speech that bypasses the higher centres of the brain entirely.

Duckspeak is an apt name for emphasis without explanation, but the differences matter more than the resemblance. Newspeak works by removing words, and it is imposed. Model prose removes nothing. Every other form of explanation remains available to anyone who wants it, and whatever drift occurs would come from convenience rather than coercion.

The appendix also carries a quieter suggestion. It is written in ordinary English and describes Newspeak in the past tense, which has led many readers to conclude that the project eventually failed. A state with full control over what people read still could not remove the capacity for unorthodox thought. That should temper any claim that patterns in chatbot prose could do so. The realistic risk is narrower than erasure: a gradual lowering of what counts as enough.

The invented example from the beginning shows what is at stake. “The problem is trust. Institutional trust. The kind cooperation depends on.” An explanation of the same point would say that cooperation depends on people expecting institutions to enforce agreements they cannot enforce themselves, and that when the expectation weakens, each party protects itself instead of contributing. The second version is longer, but its length is not what distinguishes it. It names a mechanism that a reader can check, dispute, or apply elsewhere. The first version names a subject and stops, and it stops where it does because it sounds finished.

The evidence supports the possibility that familiar language patterns influence how we express ourselves, and that the organization of prose affects what we understand. Whether sustained exposure also moves the point at which we consider an explanation complete remains unresolved. The question worth asking is not whether people will come to write like chatbots, but whether they will still notice when an explanation has ended before it began.

---

Sources

1. Wittgenstein (1953), Philosophical Investigations, trans. G. E. M. Anscombe, §§1, 87, 217.
2. Reinhart et al. (2025), Do LLMs write like humans? Variation in grammatical and rhetorical styles, PNAS. https://pmc.ncbi.nlm.nih.gov/articles/PMC11874169/
3. Bharadwaj et al. (2026), Flattery, Fluff, and Fog: Diagnosing and Mitigating Idiosyncratic Biases in Preference Models, ICLR. https://proceedings.iclr.cc/paper_files/paper/2026/hash/2096bafd3073a2224f6f0adb594068df-Abstract-Conference.html
4. Bock et al. (2007), Persistent structural priming from language comprehension to language production, Cognition. https://doi.org/10.1016/j.cognition.2006.07.003
5. Yakura et al., Empirical Evidence of Large Language Model’s Influence on Human Spoken Communication, arXiv preprint, version 4 (July 2026). https://arxiv.org/html/2409.01754v4
6. Kleijn, Pander Maat, and Sanders (2019), Comprehension Effects of Connectives Across Texts, Readers, and Coherence Relations, Discourse Processes. https://doi.org/10.1080/0163853X.2019.1605257
7. McNamara et al. (1996), Are Good Texts Always Better? Interactions of Text Coherence, Background Knowledge, and Levels of Understanding in Learning From Text, Cognition and Instruction. https://doi.org/10.1207/s1532690xci1401_1
8. Orwell (1949), Nineteen Eighty-Four, appendix: “The Principles of Newspeak.” https://orwell.ru/library/novels/1984/english/en_app