Texts from large language models often seem correct and yet are exhausting. A study by New York University and the University of Texas at Austin now offers an explanation: so-called frontier models, i.e. the most powerful language models currently available, systematically overuse linguistic constructions that make it harder for readers to process.

Among the patterns cited are unusual word combinations as well as personified objects, i.e. formulations in which things or abstract concepts act as if they were people. Such patterns are grammatically unremarkable but cost extra attention when reading.

Why the large models in particular write more clunkily

The researchers identify training on code and mathematics as the cause. There, correctness counts, not comprehensibility. A program is right or wrong; whether it reads pleasantly plays no role in the evaluation. This imprint evidently carries over into natural-language output.

What is remarkable is the finding that, according to the study, smaller models write noticeably more humanly. More model size and more training effort therefore do not automatically lead to more readable texts. In practice, this means: choosing the model is also an editorial decision, not just a technical one.

What this means for website content

On websites, this effect meets an audience that rarely reads but skims. Anyone who opens a services page, a product description or a guide decides within seconds whether to keep reading. Additional processing effort caused by awkward metaphors or overwrought imagery has an immediate effect here.

Three areas are particularly affected:

  • Comprehensibility: Technically correct but convoluted formulations miss precisely those readers who need an explanation.
  • Accessibility: Plain language, clear sentence structures and unambiguous references are part of accessible content. Unusual word combinations work against this.
  • Brand voice: When several texts come from the same model, rhythm and word choice converge. Your own tonality disappears behind a recognisable model style.

Editorial guardrails instead of doing without

The consequence is not to forgo AI support. It lies in framing its use editorially. Simple, verifiable rules in the editorial process have proven effective:

  1. Style guide as part of the prompt: Tonality, permitted sentence lengths, handling of technical terms and the company's typical formulations belong in the instructions to the model, not only in the correction loop.
  2. Maintain a negative list: Phrases that keep standing out in your own environment can be explicitly excluded and checked later during editing.
  3. Measure readability instead of estimating it: Sentence length, proportion of subordinate clauses and density of foreign words are simple indicators that can be built into the approval process.
  4. Reading aloud as a test: Texts that stumble when read aloud also stumble when read silently. This test reliably finds personified objects and awkward imagery.
  5. Make human approval mandatory: No AI text goes live without editorial sign-off. This is less a question of quality than a question of responsibility.

Implementation in the CMS

In a content management system such as TYPO3, such guardrails can be anchored technically instead of being formulated as a request to the editorial team. Approval workflows are conceivable in which AI-assisted content passes through its own status before being published. Structured content elements that limit text lengths and thus prevent overly long passages from the outset are equally helpful.

Where models are connected directly to the CMS, it is also worth making a deliberate decision about which model is responsible for which task. Summarising, rewriting and drafting place different demands. The finding that smaller models can sound more human suggests not reflexively always using the largest available model.

Assessment

The study shifts the perspective: AI texts are not just about factual accuracy, but about whether people can absorb them effortlessly. For web and software projects, this means treating readability as a measurable quality requirement and not as a matter of taste at the end of the process. Anyone using AI in content production should plan the style guide, approval and model selection as an integral part of the system architecture.

Sources