Why Is AI-Generated Language So Hard to Understand? Inside arXiv 2610.29162 on LLM Writing Style

A study posted to arXiv on 21 October 2026 finds that all nine frontier and near-frontier language models it tested use psycholinguistically harder constructions than human writers do, and proposes an explanation: AI text is increasingly optimized for information density in math, code, and reasoning rather than for successful communication. The paper, titled "Why is AI-generated language so hard to understand?" (arXiv:2610.29162, cs.CL with an cs.AI cross-list, submitted by Thomas Clark), reached 251 points and roughly 200 comments on Hacker News, where practitioners largely accepted the diagnosis and added symptoms of their own.

The paper's central claim

The widespread sense that AI-generated text has recently become harder to understand is a measurable linguistic phenomenon, not a subjective impression. The paper argues that language models overuse features that psycholinguistics already knows increase human processing difficulty, and that this overuse is a side effect of training objectives that reward information density instead of successful communication.

What the study measured

The study assembles a novel corpus of AI-generated text across diverse domains and scores it on psycholinguistic features known to correlate with human reading difficulty: object relative clauses, long syntactic dependencies, and low-frequency n-grams. Two comparison points are used: paraphrases of the same content produced by round-trip translation, and human writing.

The choice of features matters for interpreting the result. Object relative clauses, long dependencies, and rare n-grams are not stylistic preferences in this framework; they are established predictors of how long readers take and how often they misparse a sentence.

The results, by model and by feature

Compared with human writing, all nine frontier and near-frontier models tested used more inanimate subjects paired with animate verbs and more low-frequency n-grams. Most of the nine also used longer syntactic dependencies, more stacked nouns, and more reduced object relatives in conversation. The paper reports substantial variation across model families, so the behavior is not uniform across labs or model lineages.

The models also used these difficult features more often than paraphrases generated by round-trip translation, meaning the effect is not simply a byproduct of paraphrasing or of any single rewriting pipeline.

The proposed cause: information density versus communication

The paper's hypothesis is that human language is optimized for successful communication, while AI-generated text is increasingly optimized for other goals such as math, code, and reasoning, where the pressure is for information density rather than mutual understanding. Under that pressure, compressed and syntactically unusual constructions are efficient in tokens even when they are expensive for a human reader.

The authors provide a proof of concept rather than a full causal account: fine-tuning a language model under a length penalty leads to divergence from human-like linguistic behavior. In other words, rewarding shorter output alone is enough to shift grammatical choices away from what people naturally produce.

The reinforcement-learning link, in the paper's own words

The paper's conclusions, quoted in the Hacker News thread by commenter orbital-decay, tie the effect to reinforcement learning with GRPO and to verifiable rewards:

We take these results as a proof of concept that a length penalty used during reinforcement learning with GRPO can directly affect grammatical features. We note that the model was not directly instructed to use compressed or dense language, and was not directly given information about the availability of syntactic alternations. Rather, the penalty resulted in an emergent preference or shorter members of syntactic alternations. [...] We speculate that the use of verifiable rewards creates an environment where correctness is decoupled from humanlike communication, leading models to converge toward a novel, potentially less comprehensible distribution of language use.

Orbital-decay noted that the phenomenon is register-specific: it is visible in Claude's default register, and the model can still write differently when prompted.

Commenter explanations for why the style emerges

RLHF rewarded the look of good writing, not the usefulness of it

A technical writer argued that default model responses are optimized for what people think is good writing rather than what is actually good writing. The commenter dividefuel attributed this to years of RLHF tuned to raters' first impressions:

It uses lists! It has headings! It uses fun phrases like 'smoking gun'! It's long and informative! These things are relatively fast and easy to identify and select as a preference. But none of this really covers whether the response is useful for the audience. If you're not the audience, it can be hard to tell what is and isn't useful.

Commenter charlieyu1 made the complementary point that ease of comprehension is not prioritized in training at all, and doubted that frontier labs would spend money optimizing for something non-measurable and subjective.

Training corpora pull toward persuasion, jargon, and noise

Several commenters located the problem in the data. Commenter anigbrowl observed that training corpora include vast amounts of commercial speech designed to numb critical faculties and produce action, along with opinion articles written to persuade, often through rhetorical fallacies. Commenter podocarp argued that a model trained on the average of the whole internet learns from a distribution in which junk vastly outweighs good writing, and that RLHF biases further toward the average of user choices. Commenter lhwy02jsj made a related observation that training data reflects human expression that elaborates context and aims for precision, pushing generation toward complexity rather than conversational communication.

Token efficiency rewards dense noun phrases

The most concrete mechanical explanation in the thread was token economy. Commenter bassie123 argued that AI writing is optimized to pack the most information into the fewest tokens, giving the example:

(e.g., "baseline system state transition latency validation") instead of unpacking them with prepositions ("validating the latency of state transitions in the baseline system").

The same commenter speculated that readers who are good at skimming have fewer problems with AI-generated text, and that the difficulty may be concentrated among readers who depend on sequential comprehension. Commenter amelius extended the density argument past user-facing text: when AI-generated language is used internally to instruct agents and to feed reasoning loops, optimizing only for the result means the model increasingly speaks AI rather than human.

There is no mind behind the text

A cluster of comments argued that the difficulty is not a matter of word choice at all but of the absence of an author. Commenter lukev wrote that LLM text "is hard to understand because while it's perfectly good language that follows valid cognitive and informational patterns, it does not reflect the intent or mental model of another mind." Commenter Terr_ made the same argument from the reader's side: human communication is built on subconsciously simulating the mind on the other end, and LLM output fails that simulation persistently enough to be tiring, comparing it to a plastic animal that triggers mating instincts because it meets just enough of the requirements. Commenter noosphr attributed it to missing grounding: "AI generated text has no human centered underlying model of physical reality and its metaphors are tortured as a result." Commenter grey-area summarized it as "a word generator not a sense generator." Commenter Rapzid tied the same deficit to design ability, arguing that models lack cognitive empathy, which is what produces good communication in writing and in system design alike.

Specific constructions that make AI text hard to read

Transitive verbs used intransitively

Commenter cryzinger flagged a construction not named in the paper: models use transitive verbs intransitively in ways rarely seen in human writing. The example given was "The user presses a button, and the form submits." The commenter noted that valid intransitive uses of "submit" exist, but none that make sense in that context, where "submit" needs an object or should be rendered as a passive. The complaint is that ambiguity forces readers to guess, which the commenter called unacceptable in technical documentation: "if your so-called documentation requires readers/users to make educated guesses then something has gone terribly wrong."

Dense nominal style and referential ellipsis

Commenter lxe described iterating on a writing-clarity prompt and coined a name for the pattern the prompt targets: "Asyndetic parataxis with referential ellipsis and repeated rhetorical closure." The commenter noted the phrase does not fully capture the problem but comes close. Asyndetic parataxis means clauses joined without conjunctions; referential ellipsis means leaving referents implicit; repeated rhetorical closure means ending passages with the same summarizing cadence. Those three properties combined describe text that reads as dense, self-referential, and oddly rhythmic.

Ambiguity as a hedge, and garden paths as option-keeping

Commenter vikramkr offered a speculative mechanism connecting ambiguity to decoding: ambiguous references such as a bare "it" may let a model stay uncommitted until later tokens can resolve toward whatever turns out to be correct, whereas committing to a definite noun early would make a wrong answer unfixable. The same commenter questioned which parts of the style are real RL effects versus overinterpretation, noting that models do not all write alike unless they are distilling the same source, and citing a recollection that a recent Opus model was reinforcement-tuned to talk more normally and both read better and scored better than its predecessor. Commenter xg15 offered a related speculation that step-by-step reasoning under reinforcement learning pressure may encourage grammatically ambiguous sentences, because ambiguity keeps the model's options open further into the sentence.

Verbosity versus density

The thread contained a direct disagreement about whether the problem is too much compression or too little. Commenter petersumskas reported not encountering excessive concision, describing what they call "Claudish" as closer to the "become an expert in X using this one weird trick" genre: articles that go on and on and never quite get to the point, which they fix by asking for more concision. Commenter bassie123 reported the opposite problem for themselves. Both can be true because the paper's measured features are about syntactic and lexical difficulty, not raw length.

Counterpoints raised in the discussion

Commenter jasomill rejected the premise in part: "I don't find AI writing hard to understand at all, just a bit tedious," while noting that many so-called AI-isms are patterns they also edit out of their own writing for clarity, and defending em dashes as legitimate and desirable. They also observed that many of the same criticisms could be leveled at human prose.

Commenter ionetan argued the phenomenon is not uniquely artificial, pointing to politicians and academics who express themselves cleverly while saying, on close parsing, something incoherent: "AI has a way of using abstract language, sounding impressive and eloquent, but in reality it is mostly incoherent fluff." That framing implies the paper's measured features may capture a style of pretension that predates language models, even if models now produce it in greater volume.

Commenter orbital-decay's note that the effect is register-dependent is an important boundary on the claim: the same model that sounds compressed by default can be steered out of the style, which means the behavior is a default rather than a fixed capability. Commenter Self-Perfection proposed that this could be exploited deliberately, pointing to work on prompt injection as role confusion, which shows that different roles produce text with different structure and texture, and suggesting models could be tuned to keep information density in a thinking role while producing human-readable text in an assistant role.

Practical mitigations people reported using

Several commenters described steering techniques that measurably change output. Commenter cowlby reported consistently asking for "8th grade reading level," "children's story book," or ASD-STE100 Simplified Technical English, and finding that even small steering constrains language choices significantly. Commenter lxe published an iterated writing-clarity prompt to a public gist. Commenter TheSmoke reported using "unslop" and technical-writing skills from the pstack plugin collection for technical writing, system design, architecture, and code comments. Commenter unsloppable described building a tool that expands AI copy into components of meaning and then rewrites it in more ordinary language, motivated by having to comprehend large volumes of AI-generated text at work.

Open questions the thread left unresolved

Commenter sadrasabouri asked whether models trained only on natural language, such as older systems, show the same characteristics, and whether training on math, code, and reasoning helps in some ways while producing this writing style as a side effect. That question tests the paper's causal hypothesis directly and is not answered in the thread.

Commenter prmph asked why models appear to have gotten worse over time, recalling a distinct period when Claude's writing became incomprehensible and strange after having been reasonably clear before. No one in the thread produced a definitive explanation, though vikramkr's suggestion that writing style is a deliberate product decision about where to spend GPU time is one candidate.

Commenter armchairhacker raised a related failure mode that is not about syntax at all: models are bad at distinguishing the important from the mundane. Asked to improve a high-level specification, a model constantly proposes documenting arcane edge cases and implementation details. Commenter ianmarcinkowski approached the same theme from the workplace side, recounting a team conversation about colleagues who route their writing through AI, and stating a position the thread's tone largely shared: "writing is thinking," and they would rather think the thoughts. The commenter added, "I hate AI writing. So much."

Bottom line

The arXiv paper 2610.29162 packages a common complaint into testable claims: nine frontier and near-frontier models use harder psycholinguistic constructions than humans, and a length penalty applied during reinforcement learning is enough on its own to push a model away from human-like grammar. The Hacker News discussion converged on three complementing explanations for the same observation: RLHF rewards writing that looks impressive at a glance, training corpora are saturated with persuasive and commercial text, and readers are trying to model a mind that is not there. The most actionable finding across both the paper and the thread is that the style is a default rather than a hard constraint, since reading-level instructions, clarity prompts, and explicit role separation all change the output.