Harder to Read Than People, and Easier Than Last Time on 19 of 21 Counts

A paper posted to arXiv on 21 October, by Thomas Hikaru Clark, Matias Martin and Kyle Mahowald, asks the question in its title: “Why is AI-generated language so hard to understand?” Its abstract opens with the premise: “There is a widespread sense that AI-generated text has recently become harder to understand.” It was being discussed on Hacker News by 24 October.

The paper’s measurements are careful and its main finding is clear. What it measures is narrower than the premise, though, and the one piece of evidence about change over time points the other way.

What was measured

The authors count six features that psycholinguistics has linked to slower reading: low-frequency word pairs, long syntactic dependencies, object relative clauses, stacked noun compounds, verbs made from nouns, and inanimate subjects with verbs that usually take animate ones (one example the paper gives is “is_file() swallows that and returns False”). They collected text from 14 models in three settings and set it beside human writing in the same genre, and beside round-trip translations of each model’s own text into Japanese and back with a translation model, which keep the content but normalize the phrasing.

The three settings in the study and their human baselines A three-row table. Chat: 12 models, 120 conversations, 1,428 replies, compared with OpenSubtitles dialogue. Coding: the 9 frontier models, 5 repositories, 45 sessions, compared with StackExchange. Essays: 14 models, 980 model essays on 7 assignments, compared with 2,934 student essays from PERSUADE 2.0 that scored 5 or 6 out of 6. STUDY DESIGN · CLARK, MARTIN & MAHOWALD, ARXIV, 21 OCT 2026 Setting Models Model text Human baseline Chat Coding Essays 12 9 14 120 conversations, 1,428 replies 5 repositories, 45 sessions 980 essays, 7 assignments OpenSubtitles dialogue StackExchange 2,934 student essays (PERSUADE 2.0, scored 5 or 6) Each model is also compared with a round-trip translation of its own text through Japanese.
Every comparison in the study is across writers at one moment: models against people, and models against paraphrases of themselves.

The headline result holds across the design. “All nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams” than human writing, the abstract says, and most also use longer dependencies, more stacked nouns and more reduced object relatives in conversation. These are the right features to count. A 2015 study by Richard Futrell, Kyle Mahowald and Edward Gibson used parsed corpora of 37 languages and found that dependency lengths in all of them were shorter than conservative random baselines, which supports the idea that human languages keep related words close because long dependencies are harder to process. Text that lets them grow is working against that grain.

The one before-and-after pair

The premise is about change: text that “has recently become harder.” The corpus is a single snapshot, so the only way to see change in it is to compare a model with its own predecessor. The paper has one such pair, and its result is easy to miss: “Claude Opus 5.5 is lower than its predecessor Opus 5 on nearly every difficulty-linked measure in every setting (19 of 21 comparisons), with non-overlapping 95% intervals in eight.”

Opus 5.5 against Opus 5 on 21 difficulty comparisons Twenty-one squares, one per comparison. Eight dark squares: Opus 5.5 lower with non-overlapping 95 percent intervals. Eleven outlined squares: lower, with overlapping intervals. Two dashed squares: not lower. CLAUDE OPUS 5.5 VS OPUS 5 · 21 DIFFICULTY-LINKED COMPARISONS Lower, non-overlapping interval Lower, non-overlapping interval Lower, non-overlapping interval Lower, non-overlapping interval Lower, non-overlapping interval Lower, non-overlapping interval Lower, non-overlapping interval Lower, non-overlapping interval Lower, overlapping interval Lower, overlapping interval Lower, overlapping interval Lower, overlapping interval Lower, overlapping interval Lower, overlapping interval Lower, overlapping interval Lower, overlapping interval Lower, overlapping interval Lower, overlapping interval Lower, overlapping interval Not lower Not lower 8 11 2 lower, intervals apart lower, intervals overlap not lower Counts from the paper's Section 3.5. Order of squares is illustrative, not the paper's feature order.
The only model in the study that can be compared with its predecessor became easier to read, not harder, on 19 of 21 counts.

One pair is not a trend, and Opus 5.5 is still not human-like on every measure. But it is the only direct evidence on the question the title asks, and it shows one lab’s newer model moving toward people.

The mechanism, on a small model

The paper’s explanation is that models are post-trained for goals like math, code and reasoning, where there is pressure toward information density rather than toward being understood. To test it, the authors trained Qwen2.5-1.5B-Instruct with reinforcement learning on a task with two valid phrasings, adding a penalty for length. The result is clean. Without the length penalty, runs preferred the short phrasing anywhere “from 0% to 100% depending on seed.” With any penalty and no KL regularization, the preference became “over 98% short for both trained ‘that’ alternations.” The habit spread to some untrained constructions too, though “far transfer was only observed without KL penalty in the genitive alternation, one of four far-transfer alternations tested.”

Length-penalty experiment on a 1.5-billion-parameter model Preference for the short "that" alternation without KL regularization. With no length penalty, runs ranged from 0 to 100 percent depending on seed, drawn as a range line. With any length penalty, over 98 percent short, drawn as a bar. Below, far transfer: observed in 1 of 4 untrained alternations, the genitive. QWEN2.5-1.5B-INSTRUCT · GRPO · NO KL REGULARIZATION · SHORT-VARIANT PREFERENCE No length penalty Any length penalty Far transfer 0% to 100% depending on seed — paper, Section 4 Over 98% short — paper, Section 4 0–100% >98% range across seeds (top) versus every seed (bottom) Genitive: transfer observed Quantifier: not observed Passive: not observed Purpose clause: not observed 1 of 4 untrained alternations (the genitive) A proof of concept on one small model and one task, not a measurement of how frontier models were trained.
A length penalty reliably produces the short phrasing it rewards. Whether that is what happened inside the frontier models is a separate question the experiment cannot answer.

The authors call this a proof of concept, and that is the right name for it. It shows that a length pressure during reinforcement learning can push a model toward denser phrasing. It cannot show that the nine frontier models were trained under such a pressure, because their post-training is not public.

What the paper supports

I think the paper supports a precise claim and a looser one. The precise claim is that today’s frontier models write with more of the features that slow human readers than people writing in the same genre, and more than paraphrases of their own text. That is well measured. The looser claim, that AI text “has recently become harder,” is the premise the paper starts from rather than a result it measures. On the one comparison it has across time, the newer model moved the other way. A follow-up that ran the same corpus on the same model families six months apart would answer the question the title asks.

References

  1. Clark, T. H., Martin, M. B. & Mahowald, K. (2026). Why is AI-generated language so hard to understand?. arXiv:2610.29162, 21 October 2026.
  2. Futrell, R., Mahowald, K. & Gibson, E. (2015). Large-scale evidence of dependency length minimization in 37 languages. Proceedings of the National Academy of Sciences 112(33), 10336–10341.