Why AI Sentences Feel Hard to Read: Four Patterns to Fix

AI sentences often feel hard to read, and part of the reason can be counted. A new preprint found all nine frontier and near-frontier models it tested overuse sentence patterns that psycholinguistics links to processing difficulty.¹ It counted patterns, not reading times. Four of them are easy to spot and easy to undo before you save a note.

Thomas Hikaru Clark (New York University), Matias B. Martin and Kyle Mahowald (University of Texas at Austin) posted Why is AI-generated language so hard to understand? to arXiv on 21 October 2026.¹ Its first line names the complaint it tests: "There is a widespread sense that AI-generated text has recently become harder to understand."¹ Three days later its Hacker News thread had passed 150 points.²

An earlier post here covered the page-level half of this problem: why AI answers are hard to read when nothing in them ranks anything, so the eye starts to scan. This post is the sentence-level half — what happens once the eye lands on a line and has to parse it.

The study, briefly: fourteen models and a round trip through Japanese

The team collected chat, coding and essay output from 14 language models, then asked one question of every sentence: would a meaning-preserving paraphrase have phrased it differently? They answered it by translating each sentence into Japanese and back. A pattern that disappears in the round trip was a choice the model made, not something the content required.

The sample covered Claude, GPT-5.6, GLM and Kimi models, plus the open Olmo and Llama families, collected between 4 September and 11 October 2026. Chat sessions grew from ten seed prompts and "were capped at 12 turns"; coding sessions worked on five public GitHub repositories; essays answered seven PERSUADE 2.0 student assignments.¹

The control is the clever part. Translation flattens wording toward its most common form while keeping the meaning — so, in the paper's words, "a feature that becomes less frequent after the round trip is one that the model chose over an available alternative way of expressing the same content."¹

Where they could, the authors also compared against human text in a similar genre — film and TV subtitles for chat, StackExchange for coding, the student essays themselves.¹

It is a careful design. It is also a four-day-old preprint, not yet peer-reviewed.

What they actually measured, using their own sentence

The paper opens with a sentence built to fail: "Your instinct is right; the convoluted sentence you are currently reading is quietly eating its own reader-patience budget — footnote it."¹ Twenty words carry four patterns. Each has a record in research on processing difficulty, and in chat and coding each appeared more in model output than in paraphrase.

The authors list what the sentence contains: "an object relative clause, a denominal verb ("footnote"), an inanimate subject used with the verb "eat", and stacked nouns leading to a low-frequency n-gram sequence ("reader-patience budget")."¹ Their next line is the thesis — "All of these features are associated with linguistic processing difficulty."¹

The study also tracked two broader measures — rare word pairs, and the distance between words that belong together, as in "The guitar sitting in the corner because you're still "deciding" teaches you zero chords."¹

The headline result needs reading closely. "Compared with human writing, all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams, and most also use longer dependencies, more stacked nouns, and more reduced object relatives in conversation".¹ For the two all-nine features, the human yardstick is the student essays — and in chat and coding, nouns used as verbs were measured only against the paraphrases.

A single instance proves nothing. Table 2 also lists "I am writing to share my thoughts on the cell phone policy you are considering" — an ordinary sentence with a dropped that.¹ The finding is about rates.

Why models write this way: a hypothesis about density

The authors' explanation is a hypothesis, and they label it as one. Human language is shaped by the need to be understood. Recent models are trained heavily on math, code and reasoning, where they "are rewarded for getting the right answer but not necessarily for clarity."¹ That pressure, they argue, favours the shortest way to say a thing.

The abstract puts it in one sentence: "while human language is optimized for successful communication, AI-generated text is increasingly optimized for other goals like math, code, and reasoning, in which there is a pressure for information density rather than successful communication."¹

Two results point that way. Within the open Olmo family, "the effects grow from the base to the instruct to the reasoning model", while they are "absent or small for the Llama models and the Olmo base model".¹ The more post-training for problem-solving, the more the patterns showed.

The second is an experiment. The team trained a 1.5-billion-parameter model to turn structured data into English sentences, with a penalty for length. Left unconstrained, its preference for the short form of a construction became "near-categorical": "over 98% short for both trained "that" alternations" — "unlike human speakers, who choose between the variants by context."¹ That is a proof of concept — one small model, one task — not a measurement of any frontier model.

People do choose by context. T. Florian Jaeger's 2010 study of the optional that in a corpus of spontaneous speech concluded: "Information density emerges as an important predictor of speakers' preferences during production."³ Speakers add the small word when what follows is harder to predict.

An independent, peer-reviewed study points the same direction. Alex Reinhart and colleagues at Carnegie Mellon University and NJIT reported in PNAS in February 2025 that instruction-tuned models "have a distinct noun-heavy, informationally dense writing style, even when prompted to match the style of informal speech and writing."⁴ In their data, GPT-4o used nominalizations 2.1 times as often as human writers.⁴

Shorter is not the same as easier

Most plain-language advice says to cut words. For two of these four patterns, the clearer version is longer. The paper's own alternatives add words: "put the whiskey in a bottle" instead of "bottle the whiskey", and "causes an issue" where a model writes "bites".¹ The target is fewer things a reader must hold in mind, not fewer words.

The U.S. General Services Administration's plain-language guide points the other way — for a different pattern. "A hidden verb (or nominalization) is a verb converted into a noun." It goes on: "Too often, we hide verbs by turning them into nouns, making them less effective and using more words than we need."⁵

Both pieces of advice are right, because they treat opposite moves. A nominalization turns a verb into a noun — a denominal verb does the reverse. Each forces the reader to rebuild an action from a word that is wearing the wrong part of speech.

The dropped that has the strongest reading evidence behind it, though from sentences written by people, not models. Adrian Staub tracked readers' eyes in a 2010 study and found that they jumped back more often from object relative clauses than from subject relatives, and "This effect was further amplified when the relative pronoun that was omitted."⁶ One small word, restored, is cheaper than a second pass.

Why this matters for a note you will reread

A chat answer is read once, in context, with the question still on screen. A note is read weeks later by someone who has forgotten the question — usually you. Density that was tolerable in the conversation becomes a puzzle in the file — the context that resolved each stacked noun is gone.

The study hints that chat may be the worst case. Single-prompt essays showed smaller changes than chat and coding, and the authors offer a careful guess: "We speculate that certain patterns may emerge in long back-and-forth conversations (which may demand more information compression)."¹ They did not measure it.

Practitioners report something close. One Hacker News user finds that as a chat grows, "the more the model tends to come up with its own jargon to describe previous findings," and now prompts it to "summarize again without the jargon".⁷ An anecdote, not a finding.

A note has no scrollback.

How to apply this tomorrow: four rewrites before you save

These rewrites are this post's advice, not the paper's finding — the study tested no edits. They follow from the patterns it counted and the reading research it cites. Apply them when one sentence stacks several patterns at once, not to every instance.

1. Unstack the nouns. Turn a chain of three or more nouns into a phrase with a verb or a preposition. Before: "Check the team calendar sync permission settings before Friday." After: "Before Friday, check who has permission to sync the team calendar." The rewrite is longer — and has one reading.
2. Give the verb a subject that can do it. When an object eats, bites, hates or promises, name what actually happens. Before: "The import script quietly eats your timestamps." After: "The import script deletes your timestamps without warning."
3. Turn the noun back into a verb phrase. Undo the ad-hoc conversion the paper calls a denominal verb. Before: "Calendar the review and memo the team." After: "Put the review on the calendar and send the team a short memo." Do the opposite for a hidden verb: "make a decision" becomes "decide".
4. Put the that back, and keep related words close. Before: "The summary the assistant wrote yesterday, which pulls figures from three reports, two meetings and last year's budget, is wrong." After: "The summary that the assistant wrote yesterday is wrong. It pulls figures from three reports, two meetings and last year's budget." Human languages keep linked words near each other: across 37 languages, dependency lengths run shorter than random baselines.⁸
5. Ask before you copy from a long chat. Request plain sentences first, then apply the four checks to what you keep.

A request you can paste at the end of a long conversation:

Rewrite your last answer for someone reading it a month from now, without this chat.
- Use no chains of three or more nouns; spell out how they relate.
- Give every verb a subject that can actually do it; for software, use plain verbs like "deletes" or "returns".
- Do not use nouns as verbs ("calendar it", "footnote it"); use a verb phrase.
- Keep optional words like "that" when they make the structure clearer.
- Keep each subject close to its verb.
Longer is fine if it is easier to follow. Keep every fact.

The style checks that older tools automate — passive voice, nominalizations — are covered in Never Take a Word the Model Suggests. This list adds the patterns those tools never looked for.

The caveats

The study is careful, recent and narrow. It counts sentence patterns that earlier research links to difficulty; it does not test whether people read these AI outputs more slowly or understand them less. It is a v1 preprint, English-only, built on a small set of chat and coding seeds, and two of its measures rely on an automated classifier.

• Patterns, not reading times. No reading or comprehension test was run on model output. The eye-tracking evidence comes from sentences people wrote.⁶
• A small sample, by the authors' own account. "We only examine a subset of commonly used models, only conduct analyses in English, and use a limited number of conversation seeds (chat topics and code repositories)."¹
• An imperfect human baseline. Chat was compared with film and TV subtitles — spoken dialogue, which naturally carries fewer noun stacks.
• Detected by machine. Nouns used as verbs and inanimate subjects were judged by an automated classifier, not hand-coded.¹
• The cause is a hypothesis. The length-penalty result comes from one small model on one task.
• Not every reader struggles. "Nah I don't have any issue understanding, so the actual question here is why do some people find AI-generated language hard to understand?" one commenter asked.⁹ The authors concede exposure "may acclimate humans to these out-of-distribution constructions".¹
• Model choice matters; this is not a ranking. Within one family, the newer model scored lower than its predecessor on 19 of 21 difficulty-linked comparisons. The authors read that as "a quantifiable improvement" — evidence that training can push the other way.¹

One more disclosure: Kyle Mahowald co-wrote both the new preprint and the 2015 dependency-length study cited above.⁸

Frequently Asked Questions

The questions below are phrased the way people ask them in forums and search. The short version: the study counted sentence patterns linked to difficulty, the rewrites are an editor's response to that count, and none of it has yet been tested as reading time on AI output.

Why is AI generated language so hard to understand?

One measurable reason is sentence structure. A 2026 preprint from NYU and UT Austin found that all nine frontier models it tested use more inanimate subjects with verbs meant for living things and more rare word pairs than human writing, and most overuse stacked nouns and a dropped that in conversation. The authors suspect training for math and code rewards density over clarity.

Why do some people find AI-generated language hard to understand?

The study measured rates across a corpus, not a verdict on every sentence, so many AI sentences read perfectly well. Readers differ in how much density they tolerate and how well they know a topic, and the authors note that exposure may make people used to these constructions. The patterns cost most when several stack in one sentence.

How do you spot AI generated writing?

These patterns describe how AI text tends to be built, but they are a poor detector: people write stacked nouns and dropped thats too. Controlled studies find readers cannot reliably tell AI text from human text, as covered in Never Take a Word the Model Suggests. Fix sentences for the reader instead.

How do I avoid sounding like ChatGPT?

Apply the same four checks to your own drafts. Break chains of three or more nouns into phrases, give each verb a subject that can do it, replace nouns used as verbs with a verb phrase, and keep the optional that when it clarifies the structure. Read the result aloud; where you stumble, rewrite.

Does asking the AI to be concise make its answers easier to read?

Not necessarily. The paper's hypothesis is that pressure for brevity is part of the cause: in its experiment, a length penalty pushed a small model, left unconstrained, to drop the optional that almost every time. Concise and clear are different targets. Ask for fewer things a reader must hold in mind, not fewer words.

Has anyone measured how long AI text takes to read?

Not in this study. Clark, Martin and Mahowald counted sentence features that earlier research links to processing difficulty; they did not time readers or test comprehension on AI output. The eye-tracking evidence for the dropped that comes from a 2010 study of human-written sentences. Treat the claim that AI text is harder to read as plausible, not proven.

A saved sentence has lost the conversation that explained it; whatever it means, it now has to mean on its own.

mnmnote.com keeps your notes as plain Markdown on your own device, so the sentence you rewrite before saving is the one you find later.