研究总结 AI 文本难读的六种语言模式

A study by researchers at New York University and the University of Texas at Austin backs that up with data. The team had 14 language models produce text in casual conversation, coding sessions, and argumentative essays. The lineup included Claude, GPT-5.6, GLM-5, Kimi K2.5, and smaller open-source models like Llama and Olmo. In total, the researchers analyzed 120 conversations, 45 coding sessions, and 980 essays, then compared them against human-written text from movie subtitles, StackExchange, and student essays.

Six patterns that make AI text a chore to read

AI models overuse unusual word combinations that rarely show up in human writing, like "reader-patience budget," and they place related parts of a sentence far apart in ways that tax working memory. Claude Opus 5 generated "A cat who can always get up and away is a much calmer cat," where the subject and the main predicate are separated by a long clause.

Shortened references with a dropped connector are common too. Instead of "the book that the professor wrote," the models write "the book the professor wrote," saving one word but creating brief confusion about how the parts fit together. Stacked nouns like "horror movie soundtrack generator" force readers to figure out on their own which word modifies which.

The models turn nouns into verbs where humans wouldn't. Claude Fable 5 wrote "calendar it," where a human would say "put it in the calendar." They apply animate verbs to inanimate subjects, too. Claude Opus 5 produced "The whole conversation is eating its own tail," even though a conversation doesn't have a tail and can't eat. Readers can figure out what's meant, but it takes an extra beat.

Smaller models write more like humans

Llama and the Olmo base model come closest to human text. Within the Olmo family, scores rise steadily from the base model through the instruct variant to the reasoning model Olmo Think, which hits frontier-level numbers, pointing to post-training as the likely cause. Claude Opus 5.5 beats its predecessor Opus 5 in 19 out of 21 comparisons, suggesting that targeted finetuning can improve readability.

Training rewards information density over clarity

Human language balances ease of production with ease of understanding, but modern language models are trained increasingly on tasks like math, code, and logical reasoning, where correctness gets rewarded and clarity doesn't, the researchers say.

The shift from RLHF (Reinforcement Learning from Human Feedback) to RLVR (Reinforcement Learning with Verifiable Rewards) removes humans from the feedback loop entirely, the researchers argue, and models develop a kind of "dialect" that works well for compressing information and solving complex problems but poorly for human readers.

As a proof of concept, the researchers trained a smaller language model with a penalty term that rewarded shorter answers. Under certain training conditions, the model favored shorter phrasings almost without exception, regardless of context, and even generalized that preference to sentence patterns it had never encountered during training. Humans, by contrast, chose between shorter and longer variants depending on the situation.

A similar explanation comes from inside the industry. Anthropic employee Jackson Kernion, who works on Claude fine-tuning, recently explained why Opus 4.6 was the last good "writing model" from Anthropic. Optimizing newer models for math, code, and technical explanations aimed at other AI models pushed Claude toward a style that works well for AI but reads as too dense and unnatural for humans. Kernion called it an adaptation to "LLM psychology." With Opus 5.5, Anthropic pushed back for the first time, addressing sentence clarity and overloaded blocks of information, and as the study shows, successfully so.