Many readers have noticed it anecdotally: text written by AI models feels increasingly dense, stilted, and strangely exhausting to parse, even when it is grammatically flawless. A new preprint posted to arXiv puts hard linguistic data behind that intuition — and proposes an explanation for why it is happening.

The study, "Why is AI-generated language so hard to understand?", was submitted on October 21 by researchers Thomas Hikaru Clark, Matias B. Martin, and Kyle Mahowald. Drawing on a novel corpus of AI-generated text across diverse domains, the authors analyzed a set of psycholinguistic features that are well established in the literature as markers of processing difficulty for human readers — including object relative clauses, long syntactic dependencies, and low-frequency n-grams. For more context on this story, see our ongoing AI industry coverage.

What the Researchers Measured

Psycholinguistics has long known that not all sentences cost the same to read. Object relative clauses ("the senator that the reporter attacked...") take measurably longer to process than subject relatives. Long dependencies between words force readers to hold more material in memory. Rare word sequences require more effort to retrieve and integrate. These features serve as a ruler for measuring how hard a text is to actually understand — as distinct from how sophisticated it sounds.

The researchers compared AI-generated writing against two baselines: human-written text, and paraphrases produced through round-trip translation, a control designed to separate genuine model quirks from the generic flavor of formal or machine-ish prose.

That control matters. Skeptics could reasonably argue that AI text simply reads like business or technical writing — formal, impersonal, and noun-heavy — rather than exhibiting a distinct pathology. Round-trip translation produces text with a similar polished, slightly unnatural register without any language-model training behind it. If AI models still exceeded that baseline on processing-difficulty measures, the difference would have to come from something specific to how the models themselves learned to write. That, in essence, is what the study found.

What They Found

The results were consistent across the board. Compared with human writing, all nine frontier and near-frontier models examined in the study — a group that includes systems from the GPT-5.6, Claude, Llama, DeepSeek, and Qwen families — made measurably harder choices:

  • More inanimate subjects paired with animate verbs — the characteristic AI pattern of writing that abstractions "delve," "underscore," or "highlight."
  • More low-frequency n-grams, meaning unusual word sequences that human writers rarely produce.
  • Longer syntactic dependencies, with subjects and verbs separated by more intervening material.
  • More stacked nouns and, in conversational text, more reduced object relative clauses.

The study also documented substantial variation across model families — not all models are equally difficult to read — and found that AI models used these difficult constructions more often than the round-trip-translation paraphrases, indicating the effect is specific to how models are trained rather than an artifact of formal register.

That variation is quietly one of the most useful results in the paper. If reading difficulty were an unavoidable consequence of scale or capability, all frontier systems would cluster together. Instead, the spread between families suggests that post-training choices — the data mix, the reward signals, the emphasis on reasoning-heavy domains — visibly shape how a model's prose lands on a human reader. In other words, hard-to-read AI text is a design outcome, not an inevitable byproduct of intelligence.

The Information-Density Hypothesis

Why would models drift toward harder language? The authors propose and test a hypothesis: while human language evolved and is learned under pressure to communicate successfully, AI-generated text is increasingly optimized for other goals — mathematics, code, and reasoning tasks — where the pressure is toward information density rather than ease of comprehension.

A model rewarded for packing correct answers into fewer tokens learns habits that carry over into prose: compress meaning, prefer precise-but-rare wordings, embed more structure per sentence. The result is text that scores well on benchmarks while growing steadily less natural to human readers.

To test the idea, the researchers ran a proof-of-concept experiment: they fine-tuned a language model under a length penalty, forcing it to compress its output. The model's linguistic behavior diverged from human-like patterns in exactly the ways the hypothesis predicts — lending causal support to the idea that optimization pressure, not data alone, drives the shift.

Why It Matters

The findings highlight what the authors call a misalignment between the linguistic processing demands of humans and those implicitly preferred by large language models. That misalignment has practical consequences. AI assistants that produce harder-to-read text impose a hidden tax on every answer, and the effect compounds as people spend more of their reading time on model-generated drafts, summaries, and explanations.

The authors argue their framework points toward possible remedies — ways to train or tune models whose output respects human processing limits without sacrificing capability. The paper is a preprint and has not yet undergone peer review, but its central observation is already resonating: the work climbed the Hacker News front page on October 24, drawing an active debate about whether the trade-off between reasoning performance and readable prose is fundamental or fixable.

For now, the study gives readers a validated vocabulary for a familiar frustration. The text is not getting worse by accident — it is getting denser by design.

Sources: arXiv:2610.29162, Hacker News discussion

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →