Why Is AI-Written Text Hard to Understand?

What happened

Researchers at a university studied text written by AI models. Many people feel that AI writing has become harder to understand lately. The team checked a set of sentence features that are known to slow human readers down. They then compared AI writing with human writing.

Key numbers and facts

• The team tested nine frontier and near-frontier AI models. These are some of the most advanced models available.
• Compared with humans, all nine models used more inanimate subjects with animate verbs (for example, "The plan decided...") and more low-frequency word groups.
• Most models also used longer dependencies (ideas linked across many words), more stacked nouns, and more reduced object relative clauses in conversation.
• The team also found big differences between model families.

Some simple terms

Psycholinguistics is the study of how people process language in their minds. A clause is a small group of words with a subject and a verb. A relative clause adds extra information about a noun, like "the book that I read." Round-trip translation means translating text into another language and back again, which creates a paraphrase, or a rewording of the same meaning.

Why it matters

The authors think AI is trained more and more for goals like math, code, and reasoning. These tasks reward packing a lot of information into few words. Human language, by contrast, is shaped by the need to be understood easily. To test this idea, the team trained a model with a penalty for long answers. The model became less like human writing, which supports their idea. The work may help build AI that writes more clearly for people.

What to watch

• Whether AI developers test readability for people, not only accuracy.
• Whether the proposed fixes make AI text easier to read without losing its usefulness.
• Whether other researchers repeat these results with new models and languages.