Study Finds AI Homogenizing Human Expression and Thought

A growing body of research reveals that the same AI tools boosting individual productivity are quietly narrowing the spectrum of human expression, reasoning, and cultural identity — and the implications could reshape how societies think, create, and solve problems.

The Warning

On March 11, 2026, researchers at the University of Southern California published what may become one of the most consequential opinion papers of the AI era. Writing in Trends in Cognitive Sciences, computer scientist Zhivar Sourati and psychologist Morteza Dehghani synthesized evidence from more than 130 studies across linguistics, psychology, cognitive science, and computer science to deliver a stark warning: large language models are compressing the extraordinary variety of human expression into predictable, standardized patterns [1][2].

"When these differences are mediated by the same LLMs, their distinct linguistic style, perspective, and reasoning strategies become homogenized, producing standardized expressions and thoughts across users," Sourati, a PhD student at USC's Viterbi School of Engineering, told reporters [3].

The paper arrives at a moment of maximum exposure. ChatGPT alone now counts more than 800 million weekly active users — roughly 10% of the global population, according to OpenAI CEO Sam Altman [4]. Writing is the most common work task people perform with the tool, and three-quarters of all ChatGPT conversations involve practical guidance, information-seeking, or composition [4]. The pipeline through which a significant share of humanity's written output now flows has never been narrower.

The Creativity Paradox

The homogenization thesis is not merely theoretical. It rests on a growing foundation of experimental evidence that reveals a troubling paradox at the heart of generative AI.

In a landmark 2024 study published in Science Advances, researchers Anil Doshi and Oliver Hauser assigned 300 participants to write short stories under three conditions: without AI help, with a single AI-generated story idea, or with up to five AI-generated ideas to choose from. Six hundred independent evaluators then judged the results [5].

The findings were striking. Writers who used AI ideas produced stories rated as more creative, better written, and more enjoyable — particularly writers who had scored lower on baseline creativity assessments. AI acted as an equalizer, lifting the floor of creative output.

But there was a cost. AI-assisted stories were significantly more similar to each other than those written by humans alone. The researchers described the dynamic as a "social dilemma": individuals were better off with AI, but collectively, a narrower scope of novel content was produced [5].

Source: Doshi & Hauser, Science Advances (2024)

Data as of Jul 12, 2024CSV

This pattern — individual gain, collective loss — has become a recurring theme across the literature. As the USC team noted in their review, individuals using LLMs generate higher volumes of output, but groups collaborating without AI consistently produce more diverse ideas than those relying on the technology [1][3].

Writing in Someone Else's Voice

The homogenization effect is not culturally neutral. A Cornell University study presented at the ACM CHI conference in April 2025 provided some of the first direct evidence that AI writing assistants systematically push users toward Western — and specifically American — modes of expression [6].

Researchers recruited 118 participants, roughly half from India and half from the United States, and divided them into groups writing independently or with an AI assistant. The results were unambiguous: when Indian participants used the AI tool, their writing began to converge with American styles, primarily at the expense of their own cultural expression [6].

Indian users accepted 25% of AI suggestions compared to 19% for Americans, but frequently had to modify them. The AI suggested culturally inappropriate content — recommending "pizza" and "Christmas" for food and holiday prompts regardless of the user's cultural context [6].

"When Indian users use writing suggestions, they start mimicking American styles to the point of describing their own culture through a Western lens," said Dhruv Agarwal, the doctoral student who presented the findings [6].

This cultural steamrolling is baked into the technology's architecture. LLMs are trained predominantly on English-language text drawn from Western internet sources. OpenAI itself has acknowledged that ChatGPT is "skewed towards Western views" [3]. The consequence is a system that does not merely assist expression — it reshapes it, substituting the user's cultural voice with an algorithmically averaged one.

The Mind Beneath the Surface

The effects of AI homogenization extend beyond language into the structure of thought itself.

A Microsoft Research study published at CHI 2025 surveyed 319 knowledge workers and analyzed 936 real-world examples of generative AI use. The findings documented a measurable shift in how people think when working alongside AI tools [7].

Workers who expressed higher confidence in AI engaged in less critical thinking. The nature of cognition itself changed: rather than generating ideas and solving problems, workers shifted toward verifying AI outputs, integrating AI responses, and supervising AI-driven tasks. The researchers warned of "long-term reliance and diminished independent problem-solving" [7][8].

The study identified a confidence paradox: trust in AI suppressed critical engagement, while self-confidence in one's own expertise promoted it. For novices and workers in unfamiliar domains — precisely the people most likely to rely on AI — the erosion of independent thinking was most pronounced [7].

"Cognition that is unused risks becoming atrophied and unprepared when it is needed most," the researchers wrote [8].

Science Under the Same Spell

The homogenization effect is not limited to everyday writing and workplace tasks. It has reached the heart of scientific research itself.

A 2026 paper in Nature Communications Psychology documented how AI is creating a "scientific monoculture" — a self-reinforcing cycle of topical, methodological, conceptual, and linguistic convergence in academic research [9].

The numbers are striking. Scientists who adopted AI-augmented research methods published 3.02 times more papers and received 4.84 times more citations. They became research project leaders 1.37 years earlier than their peers. But collectively, their work covered 4.6% less intellectual territory than conventional scientific studies [9][10].

Source: Nature (2025) / Nature Communications Psychology (2026)

Data as of Mar 13, 2026CSV

Perhaps most alarmingly, AI-driven papers spawned 22% less engagement across natural science disciplines, suggesting reduced cross-pollination of ideas between research areas [10]. The technology that makes individual scientists more productive is simultaneously making science as a whole less exploratory and less connected.

A separate paper in Nature confirmed the pattern: AI tools expand scientists' individual impact but contract science's collective focus — the same creativity paradox observed in creative writing, now operating at the level of institutional knowledge production [10].

The Scale of Exposure

Understanding why this matters requires grasping the sheer scale at which AI is now mediating human communication.

ChatGPT's user base has grown explosively, from approximately 501 million monthly users in March 2025 to an estimated 888 million by February 2026 — a 77% increase in under a year [4]. And ChatGPT represents just one platform; competing tools from Google, Anthropic, Meta, and others add hundreds of millions more users to the ecosystem.

Source: First Page Sage / OpenAI

Data as of Mar 13, 2026CSV

Email composition alone accounts for 14.4% of ChatGPT use cases [4]. In the United States, 28% of employed adults report using ChatGPT for work [4]. When a tool with this reach encourages uniformity in expression, the cumulative effect on the texture of human communication is difficult to overstate.

The concern deepens when considering feedback loops. As AI-generated text proliferates online, it becomes training data for the next generation of models — a phenomenon researchers call "model collapse," where each iteration amplifies the homogenizing tendencies of the last. The diversity of expression that existed in pre-AI training data is progressively diluted with each cycle [1][9].

Defenders of Diversity

Not all researchers view the situation as uniformly bleak. A Nature commentary published in March 2026 explored whether some cognitive styles might be more resistant to AI's homogenizing pull, asking: "Can some brains resist?" [11].

The evidence suggests that expertise and self-confidence act as buffers. The Microsoft study found that workers with strong domain knowledge engaged more critically with AI outputs, questioning rather than accepting them [7]. Highly creative writers in the Doshi-Hauser study showed smaller gains from AI assistance — and smaller losses of originality — than their less creative counterparts [5].

The USC researchers themselves proposed solutions, arguing that AI developers should incorporate more real-world diversity into training datasets, not merely to preserve cognitive diversity, but to improve the models' own reasoning capabilities [1][2]. Morteza Dehghani, the senior author, emphasized that the goal is not to reject AI but to redesign it: models trained on more representative data would be less likely to flatten the very diversity that makes human thought valuable.

Some AI developers have begun responding. Research into "pluralistic alignment" — training models to represent multiple perspectives rather than converging on a single viewpoint — has gained traction in the AI safety community. But these efforts remain nascent, and the economic incentives of the AI industry favor optimization and consistency over the messiness of genuine diversity.

What Is Lost

The stakes of the homogenization debate extend far beyond academic concern. Cognitive diversity — the range of perspectives, reasoning strategies, and expressive styles within a population — is not a luxury. It is a survival mechanism.

Research consistently shows that diverse groups outperform homogeneous ones in problem-solving, innovation, and adaptation to novel challenges. When the USC team warns that homogenization "risks reducing humanity's collective wisdom and ability to adapt," they are drawing on decades of evidence in organizational psychology and evolutionary biology [1][2].

Consider the implications across domains. Legal briefs drafted through AI tend toward similar structures and arguments. Business strategies generated with AI assistance converge on the same frameworks. Medical case reports composed with AI tools share the same linguistic patterns. In each case, the surface quality may be high — polished, professional, competent — but the underlying diversity of approach is diminished.

The cultural dimension is equally profound. Yalda Daryani, Zhivar Sourati, and Morteza Dehghani published a companion paper arguing that AI functions as a "homogenizing engine" for culture, standardizing narratives, aesthetics, and values across societies that previously maintained distinct traditions [12]. When a technology trained primarily on English-language, Western internet content mediates expression for billions of users worldwide, the result is a slow erasure of cultural distinctiveness.

The Path Forward

The research community has not settled on solutions, but several principles are emerging.

First, transparency: AI systems should make their biases and limitations visible to users, enabling more critical engagement. Second, diversity by design: training datasets should be expanded to include underrepresented languages, cultural traditions, and modes of reasoning. Third, friction: some researchers argue that AI tools should deliberately introduce variability into their outputs, resisting the tendency toward optimization that produces sameness [1][12].

Fourth, and perhaps most fundamentally, education: users need to understand that AI outputs represent a statistical average of existing text, not an ideal form of expression. The Microsoft study found that knowledge workers who lacked skills to evaluate AI responses were most vulnerable to uncritical adoption [7]. Building AI literacy — the capacity to use these tools without being consumed by them — may be the most important intervention of all.

The challenge is immense. The AI industry is growing at extraordinary speed, with new users adopting these tools daily. The homogenizing effects documented across more than 130 studies are not theoretical risks — they are observed phenomena, already reshaping how hundreds of millions of people write, think, and create.

The question is no longer whether AI is flattening human expression. The evidence is clear that it is. The question is whether humanity will recognize the cost quickly enough to preserve the cognitive diversity that makes collective intelligence possible — before the great flattening becomes irreversible.