Artificial intelligence writing assistance can preserve our ideas while erasing part of what distinguishes the way we express them.
Researchers analyzed more than 880,000 texts from seven very different collections. They notably studied Reddit posts, scientific abstracts from arXiv, local articles, and speeches from the U.S. Congress. Their goal was to measure how the use of large language models, or LLMs, changes the diversity of writing styles.
Close-up of a backlit computer keyboard.
Credit: freestocks.org / Wikimedia Commons — CC0
To understand this, we need to distinguish the message from the way it is written. Two people can describe the same event using very different words, sentences, and levels of detail. These choices leave traces that can be associated with age, certain personality traits, or social environment.
The research team asked several AI models to rewrite existing texts. The main meaning was generally preserved. However, the differences between styles became smaller. Depending on the text collections and models tested, variation in several measures related to the construction of written works decreased by 21 to 50%.
In concrete terms, formulations that were previously very different become more similar after rewriting, even when the information conveyed remains comparable.
The researchers also tested tools capable of estimating certain characteristics of an author from their text. After the text passed through AI, their predictions changed more often. Some characteristics that were common in the data became more prominent, while others were less detectable. Some of the clues specific to the author had therefore been altered.
The researchers also tracked the evolution of large collections of texts published before and after the arrival of ChatGPT. They observed a decline in certain forms of linguistic diversity as signs of AI use increased.
This development may have consequences in fields where the way of writing itself serves as information. Language analyses are, for example, used in psychology, certain areas of medical research, or to personalize services. If AI intervenes before the analysis, it can alter clues that previously came directly from the person.
What remains to be determined is what will happen on a large scale. The authors indicate that the observed effects persist across different models, prompts, and types of text. Future research may in particular measure whether assistants designed to preserve personal style more effectively can reduce this standardization.