69% more neutral. 50% fewer pronouns. That's what happens to your writing when you let AI do the heavy lifting, according to a new peer-reviewed study accepted at ICLR, one of the top AI research conferences.
Researchers led by Natasha Jaques, a computer science professor at the University of Washington and senior research scientist at Google DeepMind, gave 100 participants a straightforward prompt: write about whether money leads to happiness. Some used AI tools freely. Others wrote on their own. The AI systems tested included Claude 3.5 Haiku, GPT-5 Mini, and Gemini 2.5 Flash.
The results were consistent across all three models, and none of them are flattering for the "just use AI" crowd.
The Numbers Tell a Clear Story
Participants who generated over 40% of their text using LLMs (large language models - the technology behind tools like ChatGPT and Claude) produced essays that were dramatically different from those written by hand. Their responses were 69% more likely to read as neutral rather than opinionated. They used 50% fewer pronouns - fewer "I think" and "my experience" statements that make writing feel human.
The AI-heavy essays also contained fewer anecdotes, less personal experience, and a noticeable shift toward impersonal, formal language. When researchers compared the AI-assisted essays against a 2021 baseline (before ChatGPT existed), the meaning and substance had diverged significantly.
Perhaps the most telling finding: heavy AI users themselves admitted their essays felt "significantly less creative and less in their own voice." They knew the writing wasn't theirs. But they reported similar satisfaction levels as lighter users anyway.
What's Actually Happening Under the Hood
The study found that LLMs made much larger edits than human editors when revising essays. Where a human editor might tighten a sentence or suggest a better word, the AI models were essentially rewriting passages wholesale - replacing the writer's voice with their own statistical average of "good writing."
This tracks with what anyone who has used these tools extensively already suspects. AI writing tools converge on a mean. They sand off the rough edges that make writing distinctive. The quirky metaphor gets replaced with a clean one. The rambling personal aside gets trimmed into a tidy paragraph. What's left is competent but generic.
The Practical Problem for Content Creators
This study puts some hard numbers behind a real tension for anyone who writes for a living. AI tools genuinely save time on drafts, outlines, and getting past blank-page paralysis. But if 40% or more of your output is AI-generated, the research suggests your content is measurably losing the qualities that make readers trust and connect with a human author.
The fix isn't to stop using AI. It's to use it differently. Treat it as a brainstorming partner or a first-draft generator, but do the actual writing yourself. The study essentially confirms what good editors have been saying for two years: AI is a useful tool for thinking, but a poor substitute for a voice.
For marketers and content creators who rely on personal authority and authentic voice to build audience trust, that distinction is worth real money.