The meaning survives while the variation shrinks
Writing assistants are often evaluated by whether they preserve the main idea and improve fluency. The study shows why that standard is incomplete. Language also carries variation in complexity, style, identity, psychological state, and social context.
When many people are rewritten toward a dominant pattern, each output may look competent while the collective record becomes less informative. The loss is distributed and easy to miss because no single sentence announces that a cultural or diagnostic signal has disappeared.
Original language should remain available as evidence
Clinics, schools, employers, and researchers should record when AI materially rewrites language and retain the original where consent and privacy rules permit. Models used for assessment should be tested on both versions to measure whether rewriting changes the resulting judgment.
Users also deserve a choice between correction and homogenization. An assistant can fix grammar without erasing dialect, cadence, cultural reference, or complexity that falls outside the model's preferred center.
Go to the source
Read the evidence behind this analysis. External links open in a new tab.
Nature Human Behaviour — Large language models and shrinking linguistic diversity
