AI: Six Tics That Fatigue Readers and Their Origins

Researchers from New York and Texas universities have evaluated 14 language models and identified six writing habits that make their texts more difficult to read. They link these biases to training methods that prioritize information density. Adjustments made to certain models, such as Claude Opus 5.5, demonstrate that it is possible to improve readability.

Training Without Human Feedback Favors an Unreadable "Dialect"

The researchers explain that recent language models are increasingly trained on tasks where accuracy is paramount, such as mathematics, coding, or logical reasoning, without considering clarity. The shift from reinforcement learning with human feedback (RLHF) to methods where rewards are automatically verifiable (RLVR) removes the human element from the feedback process. According to them, this leads models to develop a form of "dialect" that is effective for compressing information and solving complex problems but remains difficult for humans to read. To illustrate this phenomenon, the team trained a small model with a penalty on response length: under certain conditions, it almost always favored shorter formulations, even in new contexts. In contrast, humans alternated between short and long responses depending on the situation.

The Industry Describes the Same Stylistic Drift, Anthropic Responds

Jackson Kernion from Anthropic believes that Opus 4.6 was the last "good writing model" from the company. He attributes the evolution of Claude's style to optimization for mathematics, coding, and technical explanations intended for other AIs, a trend he refers to as "adaptation to the psychology of LLMs." Anthropic has indicated that it took into account the clarity of sentences and information overload with Opus 5.5. Opus 5.5 outperforms Opus 5 in 19 out of 21 readability comparisons, aligning with the goal of improving clarity.

A Protocol on 14 Models, Three Types of Tasks, and a Human Reference Corpus

Fourteen language models were utilized, including Claude, GPT-5.6, GLM-5, Kimi K2.5, as well as open-source models like Llama and Olmo. The researchers produced and examined 120 dialogues, 45 programming sessions, and 980 argumentative texts. These writings were compared to human achievements from movie subtitles, StackExchange, and student essays to identify stylistic and structural discrepancies.

Six Habits That Burden the Reading of Generated Texts

Six recurring patterns appear in the texts generated by the models. They frequently use uncommon word associations in human writing, such as "reader's patience budget." They separate syntactically related elements, increasing the reader's memory load, as seen in a sentence from Claude Opus 5 where a long parenthetical separates the subject from the predicate. Abbreviated references without connectors, such as "the book the professor wrote," are also common and can create temporary confusion. Nominal stacks, like "horror movie soundtrack generator," leave it to the reader to untangle the relationships between words. The models also transform nouns into verbs in unusual contexts, such as "calendar that" produced by Claude Fable 5, and assign action verbs to inanimate subjects, like "The entire conversation eats its own tail," generated by Claude Opus 5.

Marked Discrepancies Between Model Families and Effects of Subsequent Training

In the analyzed panel, Llama and the standard version of Olmo generate texts that are closest to those written by humans. Within the Olmo range, readability results increase from the standard model to the instructive version, and then to Olmo Think, the reasoning-oriented model, indicating the influence of additional training. For the Claude range, Opus 5.5 shows better performance than Opus 5 in 19 out of 21 scenarios, highlighting the contribution of specific refinement on readability.