LLMs Are Making Us All Think Alike — And That's the Real...

A new study from USC Dornsife reveals that LLMs are not just generating text — they are pulling human expression toward a single statistical center, erasing the stylistic fingerprints that make writing human. This is not a bug; it's the inevitable outcome of training models on the average of everything ever written.

USC Dornsife study found that LLM-generated text shows significantly lower lexical diversity than human-written text.
When humans use LLM tools to write, their own subsequent writing becomes more similar to the LLM's style — a feedback loop of homogenization.
This is not a future problem: the study analyzed 10 million Reddit comments and 100,000 academic abstracts, finding a clear trend toward convergence.
The key tension: LLMs promise to democratize writing, but they may be quietly enforcing a monoculture of expression.

What Did the USC Study Actually Find About LLMs and Writing Style?

The USC Dornsife study, published in early 2026, analyzed over 10 million Reddit comments and 100,000 academic abstracts from 2019 to 2025. The researchers used a metric called "lexical diversity" — the variety of unique words used — and found that starting in 2022, when LLMs entered mainstream use, lexical diversity began a steady decline. By 2025, AI-assisted writing showed 30% less lexical diversity than purely human-written text. The study also tracked a phenomenon they called "style convergence": writers who used LLM tools began to adopt the model's syntactic patterns and word choices, even in their own original writing. This is not about plagiarism; it is about subtle, unconscious mimicry.

Is This a New Problem or Just the Latest Chapter in Media Homogenization?

Critics will argue that media has always homogenized — from the printing press to radio to television. But this is different. Previous technologies distributed the same content to many readers; they did not change how those readers themselves produced new content. The LLM feedback loop is active: each time a person uses ChatGPT or Claude to draft an email, they absorb its statistical preferences and reproduce them. The USC researchers called this "behavioral contagion" — a term usually reserved for social media trends, now applied to syntax itself. I believe this is unprecedented because the homogenization happens at the level of cognitive process, not just consumption.

LLMs Are Making Us All Think Alike — And Thats the Real Danger

Who Benefits From a World Where Everyone Writes the Same Way?

The clear winners are the LLM providers: OpenAI, Google, Anthropic, and Meta. A homogenized writing landscape makes their models easier to train, cheaper to run, and more predictable in output. If every email, report, and blog post sounds like a smoothed-over average, the models have less noise to filter and fewer edge cases to handle. The losers are writers, educators, and anyone whose professional identity depends on a unique voice. Journalists, novelists, and marketers who rely on stylistic differentiation will find their work increasingly indistinguishable from AI-generated content. The study's data shows that this is already happening in academic abstracts: the once-distinctive styles of different disciplines are converging.

What Can Be Done to Reverse or Mitigate This Trend?

There are two paths. The first is technical: train LLMs to deliberately inject stylistic diversity, perhaps by sampling from a wider distribution of human writing or by allowing users to select a "persona" that preserves their voice. The second is cultural: educators and employers must actively teach people to write without AI crutches, treating LLM use as a supplement, not a replacement. Neither path is easy. OpenAI has no incentive to reduce the stickiness of its model's style, and schools are already struggling to police AI use. I expect the EU AI Office to take up this issue under its "fundamental rights" framework, potentially requiring LLM providers to offer diversity-preserving modes by 2027.

Dimension	Human-Only Writing	LLM-Assisted Writing
Lexical Diversity	High (baseline)	30% lower (USC study, 2026)
Stylistic Uniqueness	Individual fingerprints	Converges to LLM average
Feedback Loop	None (self-reinforcing)	Behavioral contagion (USC term)
Content Volume	Limited by human speed	Scalable, but at cost of diversity
Training Efficiency	Harder for models	Easier (less noise)
Verdict	Wins on creativity	Wins on efficiency

My thesis: LLMs are not just standardizing writing; they are standardizing thought itself, and the tech industry is ignoring this because it's profitable. In the short term, users will enjoy faster, easier writing. In the long term, we risk a monoculture of expression where the most creative ideas are filtered out because they deviate from the statistical mean. The winners are OpenAI and Google, whose models become even more entrenched as the default voice of the internet. The losers are writers, educators, and anyone who values originality. I predict that by Q3 2027, the EU AI Office will require LLM providers to include a "diversity preservation" mode that trains on a wider distribution of human writing, citing the USC study as evidence. This will be resisted by every major AI company, but it will be the first real regulatory intervention into AI's cultural impact.

By Q3 2027, the EU AI Office will mandate a "diversity preservation" mode for all LLMs sold in Europe, citing the USC study on lexical convergence.
OpenAI will launch a "Writer's Voice" feature by end of 2026 that claims to preserve user style, but it will be a marketing gimmick — the underlying model will still converge to the mean.
By 2028, academic journals will begin requiring AI-use disclosures that include a lexical diversity score, creating a new compliance industry.

The USC study proves that LLMs cause behavioral contagion in writing style — a finding with implications far beyond education.
The homogenization is profitable for AI companies because it makes their models cheaper to train and more predictable.
Regulation is the only realistic counterforce, and the EU is the most likely body to act first.
Educators must treat AI writing tools like calculators: useful for speed, but dangerous if used without understanding.
The real crisis is not that AI will replace writers, but that it will make all writers sound the same.