All nine frontier and near-frontier models in a new draft study use more inanimate subjects with animate verbs, and more low-frequency word sequences, than human writers do. “The report argues” is the pattern. The study measures English only and tests no human readers.
The authors are Thomas Hikaru Clark, Matias Martin and Kyle Mahowald (arXiv, 21 October 2026, version 1). The arXiv page lists no journal reference, so treat it as not peer-reviewed.
The short answer
- Language: English only.
- Text compared: multi-turn chat, multi-turn agentic coding, and single-reply essays from nine frontier and near-frontier models.
- Human baselines: OpenSubtitles (chat), StackExchange (coding), PERSUADE 2.0 student essays (essays), and Wikipedia.
- Main control: a round-trip translation paraphrase (English to Japanese to English with NLLB-200), which keeps the meaning but changes the wording.
- Result: all nine models use more inanimate subjects with animate verbs and more low-frequency n-grams than human writing. The effect is smaller in single-prompt essays than in chat and coding.
- Caveat: the authors say they rely on proxies to infer the cause.
Which features did the study count?
The paper measures features linked to processing difficulty in human readers:
- rare or unattested n-grams and n-gram perplexity;
- dependency length;
- object relative clauses, with and without “that”;
- stacked noun compounds;
- denominal verbs (nouns used as verbs);
- inanimate subjects with animate verbs, for example “the report argues”.
Compared with round-trip paraphrases, most frontier models overuse rare n-grams, denominal verbs, inanimate subjects and reduced object relatives. They underuse object relatives that keep the word “that”. The authors read this as a preference for dropping optional words.
Do all models do it equally?
No. Effects vary by model family. Llama and the Olmo base and instruct models change least. Olmo Think, a reasoning model, patterns with the frontier models, and the Olmo effects grow from base to instruct to reasoning.
The paper reports that Claude Opus 5.5 scores lower than Opus 5 on 19 of 21 difficulty-linked comparisons, with non-overlapping 95% intervals in eight. That is a result for those measures only, not a statement that one model writes better.
Why might it happen? The paper’s small experiment
The authors suspect pressure for information density during post-training. Their abstract says AI text is increasingly optimized for math, code and reasoning, which favors density over successful communication.
To test the idea at small scale, they fine-tuned Qwen2.5-1.5B-Instruct with GRPO under a length penalty on a sentence-generation task. Without KL regularization, the model came to prefer the shorter variant (dropping “that”) in over 98% of cases at every penalty value above zero. Some of that preference transferred to constructions it was not trained on. The paper notes it could not study models at real post-training scale, so this supports the idea without proving the cause.
What can writers and editors take from it?
The paper does not give a style guide. The suggestions below are our editing reading of its measures, not findings of the study:
- We suggest a person or thing that can act as the subject over an abstraction (“the team found” over “the analysis shows”).
- We suggest keeping the word “that” when removing it makes a clause harder to parse.
- We suggest replacing stacked nouns and nouns used as verbs with a plain phrase.
- We suggest breaking sentences where the verb sits far from its subject.
To check whether a model’s output suits your readers, test it on your own tasks as in how to run a one-afternoon AI model evaluation. Read comparison tables with five ways AI benchmark tables mislead.
What are the limits of the study?
- Only a subset of models was tested, and all analyses are in English, so nothing here applies to Arabic.
- The chat and coding conversations use a simulated user (Claude Sonnet 4.6) and a limited number of seeds.
- The measures are proxies for reading difficulty. This study did not test human readers.
What to do next
- Check the arXiv page for newer versions before you cite the study.
- Test a model on your own tasks and readers before you judge its writing.
Questions and answers about the study
Does this mean AI writing is worse?
Not by itself. The paper reports more of certain features known to slow human processing, not a quality score.
Does it apply to Arabic?
No. The authors analyzed English only.
Is the cause settled?
No. The authors propose information-density pressure in post-training and test it only in a small model.
Did the study test human readers?
No. The measures are proxies for reading difficulty.
Which model came out best?
The paper gives no overall winner. It reports Opus 5.5 below Opus 5 on 19 of 21 difficulty-linked comparisons, and Llama and the Olmo base and instruct models change least.
Terms in plain language
- n-gram: a sequence of consecutive words.
- Perplexity: how surprising a word sequence is to a language model.
- Inanimate subject with an animate verb: a thing doing a human action, as in “the report argues”.
- Post-training: the training stage after the base model, used to shape how it answers.
- Preprint: a paper shared before peer review.
Sources
Facts read on 26 October 2026. We did not run any experiment.
