Study finds six habits that make AI writing exhausting to read

By James Calloway

Researchers at New York University and the University of Texas at Austin analyzed 120 conversations, 45 coding sessions and 980 essays produced by 14 language models. Their study identified six recurring linguistic patterns that frontier models overuse, making the resulting text harder to read. The human comparison set comprised movie subtitles, StackExchange writing and student essays.

The tested lineup included Claude, GPT-5.6, GLM-5, Kimi K2.5 and smaller open-source models such as Llama and Olmo. Researchers tasked the models with casual conversation, coding sessions and argumentative essays, then compared their output with the human texts. The Decoder report does not name all 14 models.

The Decoder frames the problem through everyday use of ChatGPT and Claude. It says answers are often factually correct, yet nested sentences, odd word combinations and metaphors people rarely use make reading feel like work. The stakes are practical: a model can solve a task while still communicating poorly.

What are the six patterns that make AI writing exhausting?

The researchers grouped the recurring problems into six categories. Their examples consistently favor compact constructions, even when the compression adds decoding work for the reader.

• Unusual combinations. Models produced compounds such as “reader-patience budget.” The researchers said pairings like this rarely appear in human writing.
• Separated sentence elements. Claude Opus 5 generated “A cat who can always get up and away is a much calmer cat,” where a long clause separates the subject and main predicate. The structure taxes working memory, the limited capacity to hold information while reading.
• Dropped connectors. Models often replace “the book that the professor wrote” with “the book the professor wrote.” The shorter form saves one word but briefly obscures how the two nouns relate.
• Stacked nouns. Phrases such as “horror movie soundtrack generator” place several nouns together. Readers must infer which noun modifies which without an obvious guide.
• Nouns turned into verbs. Claude Fable 5 wrote “calendar it,” where a human would more naturally say “put it in the calendar.” The researchers presented the example as a grammatical compression that depends on context.
• Inanimate subjects given animate actions. Claude Opus 5 produced “The whole conversation is eating its own tail,” even though a conversation cannot eat or have a tail. Readers can infer the meaning, but the metaphor requires an extra interpretive step.

In plain terms, the models often pack grammatical relationships more tightly than people comfortably parse.

Why do smaller models read more like humans?

The researchers’ scores placed Llama and the Olmo base model closest to human writing. Within the Olmo family, scores rose steadily from the base model, the initial pretrained version, through the instruct variant to the reasoning model Olmo Think. Olmo Think reached what the report calls frontier-level numbers.

That progression points to post-training, the stage that shapes a base model’s behavior after initial training, as the likely cause of the stylistic gap. The supplied report does not include a controlled test that separates post-training from model size, so the explanation remains suggestive rather than definitive. The Decoder report likewise omits the underlying scores.

Claude Opus 5.5 gave a separate signal. It beat its predecessor, Claude Opus 5, in 19 of 21 comparisons, according to the researchers. They interpret that result as evidence that targeted fine-tuning, or additional training for specific behaviors, can improve readability, but the report names neither the dataset nor the procedure.

In plain terms, smaller base models can read naturally, while targeted post-training can improve the score without making the prose perfect.

Does training reward compression over clarity?

The researchers say human language balances ease of production with ease of understanding. In their account, modern language models increasingly train on math, code and logical reasoning, where correctness receives an explicit reward while clarity does not. The outcome need not be factual error; it is difficulty following how a correct answer is organized.

They also argue that movement from RLHF (Reinforcement Learning from Human Feedback) to RLVR (Reinforcement Learning with Verifiable Rewards) removes humans from the feedback loop entirely. In their explanation, models develop a compact “dialect” that works well for compressing information and solving complex problems but poorly for human readers. The researchers present that mechanism, but the supplied report does not separately verify the cause.

The researchers’ training explanation and Kernion’s account both connect denser Claude output to optimization for technical explanations aimed at other AI models. A response can therefore be easier for another model to parse and harder for a person to use without extra work. Readability is therefore a product issue, not just an academic style preference.

As 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. It even generalized that preference to sentence patterns it had never encountered during training.

Humans, by contrast, chose shorter or longer variants according to the situation. The experiment shows that a training signal can shift style, but it does not prove that length caused all six linguistic problems. The report omits the penalty weight, sample size and exact conditions, limiting how far the result generalizes.

In plain terms, a narrow reward can teach one fixed stylistic habit instead of context-sensitive judgment.

What does Anthropic’s internal explanation add?

Jackson Kernion, an Anthropic employee who works on Claude fine-tuning, recently explained why Opus 4.6 was the last good “writing model” from Anthropic. He said optimizing newer models for math, code and technical explanations aimed at other AI models pushed Claude toward prose that worked well for AI but dense and unnatural to people.

Kernion called that adaptation to “LLM psychology.” With Opus 5.5, Anthropic pushed back for the first time by addressing sentence clarity and overloaded blocks of information. The study’s 19-of-21 comparison supports the direction of that change, although it does not establish which training choices produced it.

The account provides no date for Kernion’s remarks and no Anthropic training recipe, so it remains an internal explanation rather than evidence other researchers can reproduce. Nor does a readability win establish that every reader will prefer the revised style.

Anthropic’s own account points to the same trade-off, but only a disclosed intervention would let outsiders verify it.

What does the study still leave unanswered?

The Decoder report does not identify the paper, its authors, a publication venue, a publication date or the scoring formula behind its readability results. It also omits model-by-model scores and says nothing about how the researchers adjusted the different human genres for length or format.

The human baselines, subtitles, StackExchange posts and student essays, carry distinct editing rules and audiences. That complicates any simple claim that one collection of writing defines natural prose. Those omissions do not erase the directly inspectable syntax examples, but they make the broader training explanation harder to test. A correlation between post-training and denser prose cannot prove cause without controlled comparisons.

For AI vendors, the researchers’ findings make readability a market issue as much as a research issue. If products increasingly emphasize reasoning and technical explanation, users pay the comprehension cost in time and effort even when the answer is correct. A readability benchmark tied to real preferences would offer more than another internal score.

The next test is whether a model can choose shorter or longer phrasing by context, as humans do, while retaining the compression the researchers associate with RLVR. Until then, the study offers a concrete checklist of failures and a plausible mechanism, not a complete diagnosis of awkward AI prose.