NYU and UT Austin study: frontier models write harder-to-read text

Researchers Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald posted a paper to arXiv on 21 October 2026 asking why AI-generated text feels harder to understand. According to the abstract, they measured psycholinguistic features known to correlate with human processing difficulty and found that frontier models use them more often than humans do. The Decoder, a secondary source, attributes the work to New York University and the University of Texas at Austin.

What the paper reports

The abstract says the team built a new corpus of AI text across several domains. Features included object relative clauses, long syntactic dependencies and low-frequency n-grams (rare word sequences). Two findings are stated directly:

• AI models use these difficult features more often than paraphrases produced by round-trip translation (translating text out and back).
• Compared with human writing, all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams. Most also use longer dependencies, more stacked nouns and more reduced object relatives in conversation, with substantial variation across model families.

Details from The Decoder

The Decoder reports that 14 models, including Claude, GPT-5.6, GLM-5, Kimi K2.5, Llama and Olmo, produced 120 conversations, 45 coding sessions and 980 essays, which were compared with human text from movie subtitles, StackExchange and student essays. I have not checked those counts against the paper, and they differ from the abstract's nine models.

The Decoder says smaller models such as Llama and the Olmo base model came closest to human text, and that scores rose from the Olmo base model through its instruct variant to Olmo Think, which reached frontier-level numbers. The article reads this as pointing to post-training. It also reports that Claude Opus 5.5 beat its predecessor Opus 5 in 19 of 21 comparisons.

The authors' explanation

The authors propose that human language is optimised for successful communication, while model output is increasingly optimised for goals such as math, code and reasoning, where information density is rewarded. As a proof of concept they fine-tuned a language model under a length penalty, which led it to diverge from human-like linguistic behaviour. The Decoder adds that the model favoured shorter phrasings almost without exception in some training conditions, whereas people chose depending on context. This is a hypothesis tested on one setup, not a demonstrated cause across vendors.

What the paper does not show

The abstract reports processing-difficulty features, not measured comprehension or task errors by real readers in a workplace. The sources I read do not give a reading-time penalty, so the cost to a support or operations team is unknown.

What to do

• If model output goes to customers or staff, sample 20 to 50 real responses per model version and have two colleagues mark sentences they had to reread.
• Add a plain-language instruction to the system prompt and re-run the same sample; the paper does not test prompting, so treat any gain as your own result.
• When a vendor ships a new version, re-check readability as well as accuracy before switching, since the Decoder's account suggests versions within one family can differ.