AI Editing, Measuring Voice Drift

Berkeley researcher Tom van Nuenen ran 300 personal narratives through 3 frontier models—Claude-family, ChatGPT-family, and Gemini-family—comparing 3 conditions: improve this, rewrite this, and preserve the original voice.

Measuring 13 stylometric indicators, the same-direction drift appeared across all models and all conditions.

• Contractions and first-person pronouns decreased.
• Longer words and higher vocabulary diversity appeared.
• Sentence length variation and punctuation patterns became more complex.
• Overall, the writing shifted from narration close to the speaker toward distanced narration.

The preserve voice prompt only lowered the intensity of the drift—it could not change its direction.