The internet has always been full of recognizable voices. Academic writing sounds different from Reddit. A newspaper columnist writes differently from a teenager telling a story online. Even when two people are making the same argument, their vocabulary, rhythm and sentence structure can reveal something about who they are.
AI may be quietly flattening those differences. A recent Pew Research Center analysis found signs of AI authorship or substantial AI editing in more than a third of webpages published since ChatGPT launched in November 2022. At the same time, new research published in Nature Human Behaviour suggests that the growing use of large language models is pushing human writing toward a narrower range of styles.
The concern goes beyond an internet filled with familiar words and polished sentences. Language carries traces of personality, background, values and ways of thinking. When millions of people run those words through the same handful of models, some of those traces begin to fade.
Researcher Zhivar Sourati began thinking about the problem after reading psychologist James Pennebaker’s The Secret Life of Pronouns, which explores the surprising amount of personal information contained in seemingly insignificant choices of language. As ChatGPT and other writing tools became increasingly common, Sourati wondered what would happen if those individual signals were repeatedly filtered through the same systems.
His team began by examining large collections of academic papers, news articles and Reddit posts written before and after the arrival of ChatGPT. Across all three, writing complexity began moving toward more common stylistic norms as AI tools became widespread.
Then the researchers tested the effect directly. They gave human-written texts to GPT-3.5, Gemini and Meta’s Llama 3 and asked the models to rewrite them using different prompts. The meaning generally survived. The variation did not. Depending on the model and task, differences in writing complexity fell by 21% to 50%.
In other words, AI was good at preserving what people wanted to say while subtly changing how they said it. That distinction matters because style contains information of its own.
In another part of the study, researchers used writing from people whose personality, demographics, empathy and moral values had previously been measured through psychological questionnaires. Statistical models could predict some of those characteristics from patterns in their language. Once the writing had been polished by AI, the accuracy of those predictions dropped by an average of 6%.
The machine had improved the text according to conventional standards. It had also removed some of the fingerprints of the person who wrote it.
The Rise of the Default Voice
Anyone who reads large amounts of AI-assisted writing may already recognize the phenomenon.
Certain transitions recur. Sentences settle into similar rhythms. Vocabulary becomes polished, cautious and broadly acceptable. Messy phrasing disappears. So do many of the odd choices that make one writer sound unlike another.
Pew has found signs of that shift across the web. Expressions and structural habits associated with AI-generated text have become more common since 2022, suggesting that the language of LLMs is beginning to seep into the broader texture of online writing.
There is an obvious reason people keep using them. AI is extraordinarily useful when a sentence will not cooperate, an email needs cleaning up or an argument needs structure. For people writing in a second language or working without professional editors, that assistance can be especially valuable.
The risk appears when assistance becomes substitution. LLMs are trained on enormous quantities of existing language and are exceptionally good at producing statistically plausible combinations of it. That makes them effective at generating writing that feels credible. It also means their preferred outputs tend to gravitate toward patterns that have already worked many times before.
Apply that layer to enough human writing and difference begins to shrink.
What Happens to Taste?
The implications extend beyond style. Learning to write is partly learning to think. Finding a sentence forces you to decide what you mean. Choosing between two words requires judgment. Getting stuck can expose a gap in an argument that fluent prose might otherwise conceal.
Sourati is particularly concerned about students and younger people whose reasoning habits are still developing. Someone who already knows how to build an argument can use AI to challenge, edit or sharpen it. Someone who consistently delegates the thinking itself may never develop the same internal process.
There is also a larger cultural question. Original ideas rarely arrive in perfectly standardized language. New forms of writing often begin as something awkward, local, unconventional or initially difficult to categorize. Slang evolves because communities bend language. Writers develop voices by repeatedly making choices that another person might have made differently.
A tool trained to make language smoother can easily smooth away some of the qualities that make it interesting. The irony is that as AI makes competent writing easier to produce, unmistakably human writing may become more valuable.
The strange phrase. The sentence that runs a little too long. The unexpected word. The rhythm that belongs to one writer and would sound wrong coming from anyone else.
For years, digital technology gave almost everyone a place to publish their voice. AI now gives almost everyone access to an editor. The challenge will be making sure the editor does not give everyone the same voice.
