NYU and UT Austin study finds six patterns that make AI text hard to read
Original title: Researchers find six patterns that make AI text so exhausting to read
Across 14 models, post-training and verifiable-reward training pushed text further from human writing than base models.
Researchers at New York University and the University of Texas at Austin compared output from 14 language models with human text from subtitles, StackExchange, and student essays. The set included Claude, GPT-5.6, GLM-5, Kimi K2.5, Llama, and Olmo, covering 120 conversations, 45 coding sessions, and 980 essays. They report six overused constructions, and argue that training for math, code, and verifiable rewards favors density over clarity. Llama and base Olmo stayed closest to human text; Claude Opus 5.5 beat Opus 5 in 19 of 21 comparisons.
Unconfirmed: The claim that RLVR removed humans from the feedback loop is the researchers' argument.
- The corpus was 120 conversations, 45 coding sessions, and 980 essays.
- Olmo Think reached frontier-level scores after instruct and reasoning post-training.
- A penalty for short answers made a small model shorten phrasing almost regardless of context.
- Anthropic's Jackson Kernion said math and code optimization made newer Claude denser.
- Vendors
- Anthropic
- Organizations
- New York UniversityUniversity of Texas at Austin
From the original article
3 min read at the-decoder.com
A study from New York University and the University of Texas at Austin shows that frontier language models systematically overuse linguistic patterns that are hard for humans to parse. The researchers trace the problem to training on code and math, where correctness gets rewarded but clarity doesn't. Smaller models, by contrast, write more like humans.
By Matthias Bastian · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.