Why AI Writing Feels Exhausting: Inside the 6 Patterns Ruining Prompts

If reading large language model outputs feels like a chore, you are not imagining things. A recent study by researchers at New York University and the University of Texas at Austin analyzed 120 conversations, 45 coding sessions, and 980 essays to find out why frontier AI text drains human readers.

What are the six linguistic traps in AI writing?

The research shows that frontier models systematically overuse awkward phrasing, such as stacking nouns into dense blocks like “horror movie soundtrack generator,” or separating subjects and predicates by long, distant clauses. They also drop connectors in ways that briefly halt comprehension and force inanimate objects to perform animate actions, like a conversation “eating its own tail.”

These quirks stem from model training. As labs shift focus toward math, code, and logical reasoning using verifiable rewards rather than human feedback, models develop a dense dialect optimized for machine efficiency. This information-compression trade-off ignores human clarity, taxing working memory during everyday reading tasks.

How to write around AI fatigue on Monday morning

To keep your generated drafts readable, you must actively override these machine habits in your prompts. Since smaller models and recent updates like Anthropic’s Claude Opus 5.5 show that targeted fine-tuning can improve readability, you can steer results by explicitly demanding simpler structures.

• Add strict negative constraints to your system prompts, such as banning noun stacking and metaphor misuse.
• Instruct the model to keep subjects and verbs close together in the same clause.
• Force shorter sentences by limiting word counts per paragraph during drafting sessions.
• Review output specifically for converted verbs like “calendar it” and rewrite them into plain human speech.

Is your workflow suffering from AI text fatigue?

If your team spends more time editing robotic drafts than writing from scratch, the problem is structural, not personal. Understanding that models write for other machines by default helps you adjust your prompting style to reclaim clarity without losing efficiency.