Nine AI models write text that is not easy to read

Researchers wrote a paper about the cause of the fact that AI text is not easy to read. They did tests on nine frontier AI models and compared the text with text written by persons. All nine models use more of the features that make text not easy for persons. The paper gives no number for how much worse the text is.

How it was tested

The team counted features that make text not easy for persons. Examples are long distances between related words, stacked nouns and low-frequency word groups. They used a new set of AI text from many fields.

What they found

• All nine models use more low-frequency word groups.
• All nine use more inanimate subjects with animate verbs.
• Most use long word distances and more stacked nouns in conversation.

The text of the models is also not as easy as text that went to a different language and back.

What is not known

The authors give a hypothesis: AI text is tuned for math, code and reasoning. These tasks reward dense information more than clear text. In a small test, they tuned one model with a penalty for length. The model moved away from the style of persons. The paper gives no test with readers, and no other group has done a check of it.

Insight: Training goals set the style of AI text

The paper connects hard text to a reward for dense text, not for clear text. Other news shows AI tools that read text that persons write for them: instructions in malware for code tools, and seller text for ranking agents. This reads as a move: more AI text is for tools to read, and no one is paid to make it clear for persons. Leaders must make clear text a rule in their tests.