- Author
- Thomas Hikaru Clark (New York University), Matias B. Martin and Kyle Mahowald (The University of Texas at Austin)
- Published
- 21 October 2026
- Kind
- Preprint paper with a corpus study and a small training experiment
- Original
- arxiv.org/abs/2610.29162
- Subject
- Writing with models
What it is
Three researchers test whether AI text is harder to read. They gather chat, coding and essay output from 14 models. They then test one explanation by fine-tuning a small model.
What it says
- Nine frontier and near-frontier models use more inanimate subjects and rare n-grams than human writing.
- Most also use longer dependencies, more stacked nouns and more reduced object relatives in conversation.
- Effects are smaller for single-prompt essays and differ by model. Llama and the Olmo base and instruct models change least.
- Claude Opus 5.5 shows fewer difficulty-linked features than Opus 5 in 19 of 21 comparisons.
- The authors hypothesize that rewards for math and code favor dense wording over clear communication.
- Without KL regularization, any positive length penalty made all seeds prefer the shorter variant.
What evidence it gives
The paper offers measurement on a new corpus, with human baselines and paraphrase controls. It adds a small training experiment on one 1.5B model. Results come with bootstrap intervals.
Interests and limits
Only English, some models and few conversation seeds were studied. No human readers were tested. The authors call Experiment 1 circumstantial evidence for the cause. Post-training cannot be studied at scale, so proxies stand in. Far transfer appeared in one of four constructions. Read partial: main text read, appendices not read in detail.