Why AI-written text is exhausting to read: six syntactic patterns, measured

Six measurable syntactic features — unattested n-grams, long-distance dependencies, object relative clauses, stacked noun compounds, denominal verbs, and inanimate subjects with animate verbs — account for much of why machine-written prose feels heavy. Each one taxes working memory at a different point: unpredictable word pairs stall prediction, long dependencies keep meaning unresolved, and unusual argument structure forces the reader to re-parse.

A recent preprint from researchers at NYU and UT Austin (arXiv:2610.29162) counts these features across model-generated chat, code, and essays, then compares them with genre-matched human baselines and with round-trip paraphrases of the same content. The design matters: it separates what the models had to say from how they chose to say it.

The six features

1. N-gram statistics (unattested n-grams). The study measures the rate of bigrams that never occur in a reference corpus, plus perplexity under a trigram model. Readers feel this as a phrase that predicts nothing: you stop skimming and start checking each word against the next.

2. Dependency length (long-distance dependencies). It measures the distance between syntactically linked words — subject and verb, verb and object, noun and modifier — standardized across sentence lengths. Readers feel it as a sentence whose ending keeps moving away while the beginning must stay in mind.

3. Object relative clauses. These are clauses that relativize the object ("the report that the team revised"), and the study separates clauses with an overt relativizer from reduced ones that drop "that" or "which." Readers feel the reduced form as a missing signpost: they must decide which noun fills which role, and a wrong first guess sends them back to the start.

4. Stacked noun compounds. The measure counts three or more nouns in a modifier relationship, as in "customer churn prediction pipeline." Readers feel the head noun arriving last, after a stack of qualifiers they have had to hold unresolved.

5. Denominal verbs. These are verbs formed directly from nouns, such as "to bottle" or "to footnote." Readers feel a part-of-speech mismatch: the word looks stored as a noun but arrives in a verb slot, so comprehension pauses to reclassify it.

6. Inanimate subjects. The study counts inanimate subjects paired with active verbs that normally take animate agents. Readers feel a beat of interpretation — "the config decides," "the error swallows the record" — and spend it deciding whether this is metaphor, shorthand, or a mistake.

How the measurement works

The corpus spans three settings: chat sessions of up to 12 turns, coding sessions of 11 turns plus a fixed closing message, and single-turn essays written to seven prompts. Nine frontier or near-frontier models took part, alongside smaller open models at different training stages. Human baselines were matched by genre: subtitles for chat, StackExchange for code, and student essays for the essay condition. A separate reference corpus supplies the n-gram counts.

Dependency structure comes from automatic parses. Dependency length is standardized within sentence-length bins, so the effect is not just "long sentences have long dependencies." N-gram difficulty is measured two ways: the rate of unattested pairs against the reference corpus and trigram-model perplexity.

The key control is round-trip translation. Each sentence is translated independently from English to Japanese and back with NLLB-200 3.3B. The content stays fixed while the wording changes. If a feature becomes rarer after the round trip, the model had an easier alternative available and chose the harder one; the difficulty is a stylistic choice rather than a demand of the content. Confidence intervals come from bootstrapping over whole conversations.

A second experiment fine-tunes a 1.5B model with reinforcement learning under a length penalty. As the penalty grows, the model shifts toward shorter members of syntactic alternations, including constructions it was never explicitly trained on. That is a proof of concept for the compression hypothesis, not a causal identification.

The main result: all nine frontier and near-frontier models used more inanimate subjects with animate verbs and more rare n-grams than the human baselines. Most also used longer dependencies, more stacked nouns, and more reduced object relatives in conversation. Effects were smaller in the essay setting, and they varied substantially across model families.

What the study does not show

• It is English-only. Every analysis is on English text. None of it covers Chinese, where topic-comment structure, measure words, and verb compounding change what counts as a long dependency or a noun stack. The specific numbers should not be transplanted to Chinese prose.
• The sample is limited. A subset of models, a limited number of conversation seeds and code repositories. It is a structured sample, not a census of model output.
• The comparison is relative, not absolute. Round-trip translation flattens style. The result shows that a model preferred the harder variant when an easier one existed; it does not quantify how much harder the final text is to read.
• There is no reading-time experiment on these texts. The processing costs are inherited from psycholinguistic work on the same constructions, not measured by timing readers on model output.
• The causal story is a hypothesis. Information density as a training pressure is supported by a small intervention, not proved. Genre also matters: essay effects were weaker than chat effects.

Five checks you can run on a draft

1. Close the dependency. Keep subject and verb close, and move qualifying clauses after the verb. "The policy, after months of review by a committee that included two former regulators, changed" becomes "The policy changed after a months-long review by a committee that included two former regulators."

2. Turn nouns back into verbs. "Make a decision" becomes "decide"; "provide an explanation" becomes "explain"; "perform an analysis" becomes "analyze." When a noun carries the action, the sentence spends an extra word on ceremony.

3. Restore the connective. When two clauses sit side by side, write the relation: "because," "so," "but," "for example." Readers should not have to infer the logic from adjacency alone.

4. Name a subject that can act. Replace abstractions in the subject slot with the thing that does the work. "The configuration causes failures" becomes "A missing timeout causes the retry loop to fail."

5. Cap noun stacks at two. Unpack three-noun chains with a preposition or a verb: "customer churn prediction pipeline" becomes "a pipeline that predicts customer churn."

These checks do not replace editing judgment. They give you five specific places to look, each tied to a construction with a known processing cost. Run them in order — dependencies first, then verbs, connectives, subjects, and noun stacks — and a pass over a thousand words usually takes a few minutes.

The useful conclusion is diagnostic rather than tribal: difficulty can be counted, and counted features can be revised. Whether a paragraph came from a person or a machine, the same six patterns predict where a reader will slow down.

Based on arXiv:2610.29162 (preprint).

If you want to see how I apply these checks to real drafts, I keep worked examples here: https://hhfa008.github.io/personal/cases.html