Voice Under Revision: Large Language Models and the Normalization of Personal Narrative
This study demonstrates that large language models exert a "directional pull" toward stylistic normalization, altering personal narratives by reducing idiosyncratic markers like pronouns and contractions while increasing abstraction, thereby making texts more polished but less uniquely situated.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you have a collection of handwritten letters from your grandparents. They are full of unique quirks: messy handwriting, little inside jokes, specific slang, and even some spelling mistakes that make them feel real. Now, imagine you ran all those letters through a high-tech "Auto-Correct" machine that promised to make them "better."
The machine would fix the spelling and make the sentences flow perfectly. But when you read them back, the magic is gone. The letters no longer sound like your grandparents; they sound like a polite, slightly boring textbook.
This is exactly what this research paper is about.
The Core Discovery: The "Smoothie" Effect
The researcher, Tom van Nuenen, wanted to see what happens when we use powerful AI (like ChatGPT or Claude) to "improve" personal stories. He took 300 real, raw stories—people talking about their lives, their heartbreaks, and their joys—and asked three different AI models to rewrite them.
He found that the AI acts like a giant industrial blender.
No matter how much "flavor" or "personality" the original story had, the AI tends to blend everything into the same smooth, consistent texture. In the science world, he calls this "Normalization."
The Three Big Changes
The paper points out three specific ways the AI "smooths out" our humanity:
The "Polite Professional" Filter (Register Shift):
The AI hates informality. It takes "I can't" and turns it into "I cannot." It takes messy, breathless sentences and turns them into long, perfectly punctuated ones. It’s like taking a person wearing a comfortable, stained t-shirt and forcing them into a stiff, starched business suit. They look "better," but they don't look like themselves anymore.The "Loss of Self" (Stylometric Convergence):
Every human has a "linguistic fingerprint"—the specific way we use words and punctuation. The study found that when AI rewrites stories, it wipes away those fingerprints. If you showed a computer the rewritten stories, it wouldn't be able to tell which person wrote which story. They all start to look like they were written by the same "Generic AI Person."The "Wise Observer" Problem (Narrative Stance):
This is the most poetic finding. When people tell real stories, they are usually in the moment (e.g., "I walked into the room and felt scared"). The AI tends to move the narrator away from the action and makes them sound like a philosopher looking back from a distance (e.g., "Looking back, the atmosphere of the room was one of palpable tension"). It turns a lived experience into a summarized report.
Can we stop it? (The "Voice-Preserving" Prompt)
The researcher tried to trick the AI by saying, "Please rewrite this, but keep the author's unique voice!"
It helped a little bit, but not much. It’s like telling a chef, "Make this meal fancy, but keep it tasting like home cooking." The chef might use better ingredients, but the dish is still going to end up looking like a high-end restaurant meal. The AI is simply "programmed" to be polished, and that instinct is very hard to turn off.
Why does this matter?
This isn't just about writing better emails. It matters for history and science.
If future historians try to study how people felt in the 2020s, but all our digital records have been "cleaned up" and "smoothed out" by AI, they won't be studying us. They will be studying the "smoothie" the AI made. We risk losing the beautiful, messy, and unique fingerprints that make human history actually human.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.