Study Reveals 6 AI Writing Habits That Exhaust Readers and How Fine-Tuning Fixes Them
Research from NYU and UT Austin shows how training AI for math and code makes output dense and tiring, and how models like Claude Opus 5.5 are shifting back toward human readability.
Recent research from New York University and the University of Texas at Austin provides empirical evidence for why AI-generated text often feels exhausting to read. In a study evaluating 14 language models across 120 conversational transcripts, 45 programming sessions, and 980 argumentative essays, researchers compared synthetic outputs against human baseline text drawn from student essays, movie subtitles, and StackExchange discussions. The evaluated lineup included prominent frontier systems such as Claude Opus 5, GPT-5.6, GLM-5, and Kimi K2.5, alongside open-source architectures like Llama and Olmo. The investigation identified six specific structural patterns that AI models systematically overuse, placing an unnecessary cognitive burden on human working memory.
Six linguistic patterns that exhaust working memory
The researchers highlighted six distinct structural habits that make model-generated text frustrating to navigate:
- Unusual phrasing: Models frequently generate rare word pairings, such as "reader-patience budget," that seldom appear in natural human communication.
- Disjointed sentence structure: Key sentence components are separated by lengthy clauses. For instance, Claude Opus 5 generated the sentence, "A cat who can always get up and away is a much calmer cat," placing a long clause between the subject and its main predicate.
- Dropped connectors: AI models regularly omit relative pronouns, writing "the book the professor wrote" instead of "the book that the professor wrote," creating momentary confusion about how sentence parts fit together.
- Stacked nouns: Chains of unlinked nouns like "horror movie soundtrack generator" force readers to decipher modifier relationships without explicit guidance.
- Nouns converted into verbs: AI models turn nouns into verbs in contexts where humans would not, such as Claude Fable 5 outputting "calendar it" instead of "put it in the calendar."
- Animate verbs for inanimate subjects: Models apply living actions to non-living concepts, with Claude Opus 5 writing that a conversation was "eating its own tail."
The shift from human feedback to information density
The study attributes these tiring quirks to fundamental shifts in post-training techniques. Natural human language balances ease of expression with ease of comprehension. However, modern model training increasingly prioritizes complex reasoning, mathematical problem-solving, and code execution, where factual correctness is rewarded while stylistic clarity is ignored. The transition from Reinforcement Learning from Human Feedback (RLHF) to Reinforcement Learning with Verifiable Rewards (RLVR) removes human evaluators from the feedback loop entirely. Consequently, models develop synthetic dialects optimized for maximum information compression rather than human readability.
Base models like Llama and the standard Olmo base model produced text closest to human writing styles. Within model families, text drifted further from human norms as models moved from base to instruct variants, reaching maximum divergence in specialized reasoning variants like Olmo Think. Industry insiders confirm this trend: Anthropic fine-tuning engineer Jackson Kernion noted that Claude Opus 4.6 was the company's last traditional "writing model" before fine-tuning shifted heavily toward machine-oriented technical tasks and "LLM psychology." However, targeted post-training can successfully reverse this drift. Claude Opus 5.5 surpassed Opus 5 in 19 out of 21 readability comparisons after Anthropic specifically addressed sentence clarity and overloaded information blocks.
What it means for developers
For software engineers integrating LLMs into applications, these findings highlight a crucial tradeoff between technical reasoning and conversational user experience. High-reasoning models optimized via RLVR excel at formal logic, debugging, and complex calculations, but using them directly for customer support, documentation, or natural dialogue can result in dense, taxing prose for end users.
Engineering teams must carefully select model checkpoints based on the explicit task or apply system prompts and post-processing that reduce noun stacking and complex clause separation. When building applications that demand a balance of technical capability, concise copy, and cost efficiency, developers can test top AI models cheaply through a single unified API at https://apixoai.online. Evaluating different model families side-by-side allows teams to pinpoint the exact threshold where reasoning power aligns with natural human readability.
Source: Researchers find six patterns that make AI text so exhausting to read — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
