The same tools that were supposed to democratize writing are now homogenizing it into a single, increasingly bland voice.
The Summary
- More than a third of web pages published since ChatGPT launched in November 2022 were AI-authored, and new research in Nature Human Behavior shows LLMs are pushing human writing toward narrower linguistic norms
- When GPT-3.5, Gemini, and Llama 3 rewrote human texts, they reduced variation in writing complexity by 21% to 50% while preserving meaning
- Analysis of 80,000 academic papers, 400,000 news articles, and 300,000 Reddit posts shows writing complexity converged toward common norms after ChatGPT's introduction
- The linguistic signals that reveal identity, background, and values are being stripped away at scale
The Signal
Researcher Zhivar Sourati started with a question that should concern anyone building or using AI writing tools: what happens when everyone's text runs through the same models? The answer, based on his analysis of over 780,000 pieces of writing, is that we get less human diversity and more machine-mediated sameness.
The numbers are stark. Across academic papers, news articles, and Reddit posts written before and after November 2022, variation in writing complexity increasingly converged toward common stylistic norms. Not in one domain. In all three. The pattern held whether someone was publishing research, reporting news, or sharing stories on social media.
"When everyone's text goes through the same models, the artifacts that reveal who we are start to disappear."
This isn't just about style. It's about identity signals. James Pennebaker's research showed that personal values, backgrounds, and beliefs shape how people communicate the same ideas using different words. Your pronoun use, sentence structure, and word choices leak information about who you are. LLMs, trained to optimize for clarity and coherence, sand those edges off.
The controlled experiment makes this concrete:
- Three different models (GPT-3.5, Gemini, Llama 3) all exhibited the same behavior
- Meaning was preserved, but stylistic variation dropped by 21% to 50%
- The effect was consistent regardless of prompt variation
What's being lost here isn't quality. AI often makes writing clearer, more concise, more accessible. What's being lost is the fingerprint. The tell. The distinctive markers that make your writing yours.
Here's where this gets tricky for the agent economy. If AI agents are going to represent us, negotiate for us, communicate on our behalf, they need to carry some version of our linguistic identity. But if the same base models are powering all of them, we're heading toward a world where every agent sounds like every other agent. Efficient, clear, and utterly interchangeable.
The convergence Sourati documented isn't a bug. It's what happens when you optimize language through a single statistical lens at massive scale. LLMs learn what "good" writing looks like from their training data, then reproduce that pattern. When millions of people use the same tool to polish their prose, you don't get millions of unique voices. You get millions of variations on the model's voice.
The Implication
If you're building AI writing tools, this research suggests you need mechanisms that preserve user voice, not just improve clarity. Maybe that means fine-tuning models on individual writing samples. Maybe it means giving users explicit controls over formality, complexity, and style dimensions. The technical challenge is real: how do you help someone write better while keeping them recognizable?
For everyone else, the message is simpler. Use AI as a tool, not a replacement. Let it catch typos, tighten arguments, suggest better structures. But don't let it erase the parts of your writing that sound like you. The rough edges, the unusual word choices, the sentence rhythms that are yours alone. Those aren't flaws to be smoothed away. They're signals. And in a world of increasing linguistic homogeneity, distinctive signals become more valuable, not less.