Language is not just a vehicle for meaning. It is a fingerprint. The specific words a person chooses, the complexity of their sentences, the rhythm of their reasoning: these carry information about who they are, where they come from, what they value. A growing body of research now suggests that large language models, when used to assist or rewrite human text, quietly sand down those fingerprints.
This is what most coverage of AI writing tools misses. The conversation tends to focus on quality, speed, or accuracy. The question of what gets lost in the process receives far less attention.
The Convergence Nobody Measured
Researcher Zhivar Sourati began asking this question after reading James Pennebaker’s book The Secret Life of Pronouns, which explores how personal values, backgrounds, and beliefs shape the specific words people use to express the same ideas. As large language models became more widely adopted, Sourati started wondering whether routing text through these systems might compress that natural variation.
To find out, his team conducted a time-series analysis drawing on approximately 80,000 academic papers, 400,000 news articles, and 300,000 Reddit stories, comparing writing produced before and after ChatGPT’s November 2022 launch. The finding was consistent across all three datasets: after the adoption of AI writing tools, variation in writing complexity converged toward common stylistic norms. The trend held regardless of the domain or the audience.
This matters because the convergence was not the result of people consciously choosing to write more simply or more uniformly. It emerged as a byproduct of passing text through systems trained on the same underlying data, optimized for the same notion of clarity and correctness.
Identity Signals, Quietly Erased
Sourati’s team then ran a more controlled experiment. They had GPT-3.5, Gemini, and Meta’s Llama 3 rewrite human-authored texts using a range of prompts. The models preserved meaning well. What they reduced was variation: writing complexity dropped by between 21% and 50% depending on the model and the prompt.
The third stage of the research pushed further. The team worked with texts written by individuals whose personal traits had already been identified through psychological questionnaires, covering demographics, personality, empathy, and moral values. After those texts were processed through a large language model, statistical models were used to predict the authors’ traits based on linguistic features. Predictive accuracy declined by an average of 6%.
That number deserves a moment of attention. A 6% average decline means the text, after AI polishing, carried measurably less information about the person who wrote it. The linguistic signals that connect words to identity had been partially erased. The results were not uniform: different personal traits were affected in different ways, which suggests the homogenizing effect is not random noise but a structured alteration of how identity is encoded in language.
According to the Pew Research Center, more than a third of all internet web pages published since ChatGPT’s launch were authored by AI. The scale of that shift means the homogenization Sourati documented is not a laboratory curiosity. It is already reshaping the texture of written communication at a population level.
Why Cognitive Diversity Is the Real Stakes
Sourati’s concern extends beyond linguistics. His argument is that linguistic diversity and cognitive diversity are connected. When everyone’s writing is filtered through systems that converge on a shared notion of what sounds credible or socially appropriate, the range of perspectives that can be expressed clearly and distinctly begins to narrow.
This is the deeper implication. It is not simply that AI-assisted writing sounds similar. It is that the tools have the capacity to define what a legitimate, well-expressed perspective looks like, and to quietly reshape text toward that definition. Sourati describes this as the ability to rewrite what counts as a credible way of having a perspective.
The population Sourati identifies as most vulnerable is students: people who are still developing their reasoning abilities and who may reach for AI assistance before they have built the cognitive habits that make independent thought possible. The risk is not that they will write worse. It is that they may never fully develop the internal process that writing, in its unassisted form, trains.
His proposed resolution is not to abandon these tools but to sequence their use carefully. People who already know how to reason through a problem can use large language models to handle specific tasks without surrendering the underlying process. The danger arises when the tool substitutes for the development of that capacity rather than extending it.
This is precisely where NAVION’s framing of AI as an augmentation layer, not a replacement for human judgment, becomes structurally important. A tool that handles volume and formatting while the human retains the reasoning is a different proposition from a tool that generates the reasoning itself.
In Short
Large language models, when used to rewrite or polish human text, measurably reduce stylistic variation and erode the linguistic signals that carry information about individual identity. Research by Zhivar Sourati, published in Nature Human Behavior, found that AI rewrites reduced variation in writing complexity by 21% to 50%, and reduced the predictive accuracy of identity-linked traits by an average of 6%. The concern is not aesthetic. It is that homogenized language reflects and reinforces homogenized thinking, and that the people most at risk are those who adopt these tools before they have learned to think independently without them.
Based on reporting from Fast Company - Tech.