The Flattening of Language
For all the emphasis on large in large language models, the diversity of their outputs is turning out to be remarkably small, and it may be dragging human expression down with it. A new study examining the widespread adoption of AI writing tools has found measurable evidence that AI-assisted text is converging toward a narrower range of styles, vocabularies, and rhetorical patterns than human-only writing produces.
The findings add empirical weight to a concern that linguists, educators, and cultural commentators have raised since generative AI tools became mainstream: that outsourcing writing to AI systems trained to produce the most statistically probable text will gradually erode the richness and diversity of human expression.
Measuring the Homogenization Effect
The research team analyzed millions of text samples across multiple domains, including academic papers, business communications, social media posts, creative writing, and journalism, comparing pieces written before and after the widespread adoption of AI writing assistants.
The results revealed consistent patterns of convergence. AI-assisted texts showed reduced lexical diversity, using a smaller range of distinct words relative to total word count. Sentence structures became more uniform, gravitating toward a middle range of lengths and complexity while avoiding both the very simple and the elaborately complex constructions that characterize natural human writing.
Most strikingly, AI-assisted texts from different authors, cultures, and languages showed greater similarity to each other than comparable human-only texts. The AI tools appeared to be acting as a stylistic averaging function, smoothing out the individual quirks, cultural influences, and personal voice that make human writing distinctive.
The Mechanism of Convergence
The homogenization occurs through a straightforward mechanism: large language models generate text by predicting the most probable next word based on patterns in their training data. This process inherently favors common patterns over rare ones, mainstream expressions over idiosyncratic ones, and conventional structures over experimental ones.
When humans use these tools as writing assistants, accepting suggested completions or using AI to draft initial versions, they incorporate this statistical averaging into their own output. Over time, as AI-assisted writing becomes the norm, the baseline of what normal writing looks like shifts toward the AI's preferred patterns.
The effect is compounded by a feedback loop. As more AI-generated text appears online, it becomes training data for future AI models. These newer models learn from an increasingly homogenized corpus, producing even more uniform outputs. The researchers describe this as a narrowing spiral.
Cultural and Intellectual Consequences
Language is not merely a vehicle for transmitting information. It shapes how people think, what concepts they can express, and how they understand the world. Different writing styles reflect different ways of processing experience. When those styles converge, the underlying diversity of thought may converge as well.
The research found particular concerns in academic writing, where disciplinary jargon and specialized rhetorical conventions serve important epistemic functions. AI tools tend to smooth these disciplinary differences, producing text that reads more like general-purpose prose than specialized discourse.
Creative writing showed the most dramatic effects. AI-assisted fiction and poetry exhibited significantly less experimentation with form, voice, and narrative structure than comparable human-only work.
The Multilingual Dimension
The homogenization effect is especially pronounced across languages. AI writing tools, predominantly trained on English-language data, tend to impose English rhetorical patterns even when generating text in other languages. Writers using AI assistance in Mandarin, Arabic, Spanish, and other languages produced text measurably more similar to English-language patterns than text written without AI assistance.
This represents a form of linguistic and cultural imperialism that operates through algorithmic optimization rather than political power. The rhetorical traditions and stylistic conventions that distinguish different literary traditions are being quietly eroded by tools that have internalized English-dominant patterns as the default.
Language preservation advocates have flagged this as a serious concern for smaller languages and literary traditions that lack large digital corpora.
Pushback and Solutions
Proponents of AI writing tools argue that clearer, more standardized prose serves communication better than idiosyncratic writing. In professional contexts, consistency and clarity are valued over individual style.
However, the researchers note that the choice between diversity and standardization should be conscious, not an accidental side effect of algorithmic design. They propose several interventions: AI tools with diversity modes that deliberately introduce variation, training data curation that prioritizes stylistic diversity, and transparency features that highlight where AI patterns are influencing a user's text.
The research ultimately raises a question that goes beyond technology: in an age when algorithms increasingly mediate human expression, who decides what counts as good writing? If the answer is a statistical model optimizing for the average, the unique voices and traditions that make human language rich may be the cost.
This article is based on reporting by Gizmodo. Read the original article.
