The “AI voice” has a measurable cost #
AI chatbots often produce prose that feels…off. A recent study from NYU and UT Austin researchers quantified this intuition, measuring sentence structures associated with increased human processing effort in outputs from 14 prominent AI models. This analysis reveals a measurable cost to the ubiquitous “AI voice.”
Researchers employed spaCy, an open-source Python library, to parse the grammatical dependencies across diverse AI-generated texts. They collected conversations, essays, and coding sessions from models including Claude Opus 5/5.5, the GPT-5.6 family, GLM-5, Kimi K2.5, Llama, and OLMo. This allowed them to identify patterns in dependency length and structure.
They then assessed these patterns against findings from psycholinguists, who have long studied how specific linguistic features affect human reading time and comprehension. These features include:
- Stacked nouns (e.g., "reader-patience budget")
- Long syntactic dependencies
- Omitted optional words like "that"
- Animate verbs with inanimate subjects
While the study meticulously measures the prevalence of these linguistic features, it’s crucial to clarify the stakes. The analysis identifies characteristics linked to slower or more effortful reading; it does not directly prove every reader struggles with every single AI-generated sentence. The findings indicate a systematic bias towards complexity, not universal incomprehension.
Six tiny habits make readers work harder #
Psycholinguists identified six specific patterns that make human brains work harder. AI models disproportionately employ these, forcing readers to parse sentences like, “The convoluted sentence you are currently reading is quietly eating its own reader-patience budget.” This example combines a stacked noun ("reader-patience budget") with an inanimate subject performing an animate action ("sentence...eating").
Other common patterns include long gaps between grammatically linked words, rare word pairings, and reduced clauses—sentences where an optional "that" is dropped (e.g., "the book the professor wrote" instead of "the book that the professor wrote"). Denominal verbs, which turn nouns into verbs (e.g., "calendar it"), also increase processing effort.
To confirm that AI models chose these less common phrasings, researchers performed a Japanese-to-English round-trip translation using Meta’s NLLB-200. This process preserves meaning but rewrites sentences into more conventional structures. If a pattern disappeared after the translation, it indicated the original AI output was unnecessarily convoluted.
For almost every frontier model, these difficult patterns vanished post-translation, proving the models gravitate toward harder-to-read constructions even when simpler alternatives are available. Smaller models like Llama showed minimal adoption of these habits, but the OLMo family exhibited a stronger tendency with each training stage.
The training incentive hiding in plain sight #
Proposed mechanisms for this AI "voice" point to post-training incentives. Models undergoing further training on tasks like math, coding, and reasoning are often rewarded solely for correct answers, with little emphasis on the explanation's readability. Furthermore, strict token budgets can inadvertently favor compressed, dense phrasing, as models learn to convey maximum information with minimum word count.
Researchers tested this hypothesis directly. They trained a Qwen 2.5 1.5B parameter model using GRPO to convert JSON event descriptions into English sentences, introducing a small penalty for each extra token. Even this minor penalty drove the model to consistently choose shorter constructions, often dropping the optional "that" over 98% of the time in some runs. This behavior emerged even in sentence types the model was not explicitly trained on, demonstrating how token economy can shape syntactic preferences.
Model behavior varied significantly. The study found little of this "denser" tendency in Llama outputs. In contrast, OLMo models showed progressively denser outputs across their training stages, with the reasoning-tuned version exhibiting the strongest preference for these cognitively taxing structures. These findings offer strong evidence that post-training objectives, particularly those rewarding conciseness over clarity, contribute to the AI voice problem. For further reading on the specifics of the study, see Why is AI-generated language so hard to understand?.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
Clear writing is a choice labs can reward #
Encouragingly, model behavior can shift: Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 measured difficulty indicators. This suggests that labs can actively reward clearer communication, demonstrating that convoluted prose isn't an immutable trait.
To steer AI toward more accessible output, prompt engineering should explicitly request plain English, fewer stacked nouns, and the use of explicit connectors where helpful. Prioritize short distances between grammatically linked words, such as subjects and verbs, to ease processing.
Crucially, clarity must be evaluated independently of factual correctness during model training and fine-tuning. This ensures readability doesn't become a secondary concern to accuracy, which, as the NYU and UT Austin researchers highlight, can lead to the "reader-patience budget" problem.
In high-stakes environments—like code reviews, medical diagnoses, or financial reports—dense AI-generated text places an undue burden on human reviewers. A dedicated clarity-focused rewrite pass can mitigate this, but always ensure the meaning is preserved and accuracy is re-verified. Clear writing is a design choice, not an accident.
Frequently Asked Questions #
Why do AI models produce hard-to-read sentences?
The study suggests post-training incentives can reward concise, information-dense answers without adequately rewarding clarity.
What sentence patterns make AI writing harder to read?
Examples include stacked nouns, long gaps between related words, unusual word pairings, omitted connectors, and inanimate subjects doing animate actions.
Did the research prove that every AI model writes this way?
No. The researchers found differences across models, tasks, and training stages; some tested models showed weaker tendencies than others.
How can teams make AI writing clearer?
Prompt for plain language, ask models to keep related words close, and use a separate editing pass focused on readability.