AI Homogenization: How Large Language Models Are Flattening Linguistic Diversity and Personal Expression

Bland new world: is AI making us all think the same?

Zhivar Sourati often gets déjà vu while perusing recent research in his field of computer science. "I read papers and I'm like, 'I've seen this paper before,'" he says. Everything's looking kind of the same, shorn of individual quirks, says Sourati, who is a PhD student at the University of Southern California in Los Angeles. Sourati suspects artificial intelligence is to blame. Because people are increasingly relying on large language models (LLMs), their writing is growing more similar, he suggests. Now, Sourati and other researchers are putting their anecdotal observations to the test by exploring how generative AI (genAI) -- systems that create text, images and other output -- might be homogenizing both culture and cognition. In a March paper, he and his co-authors noted that this phenomenon is similar to the concept of 'McDonaldization', invoking how characteristics of the fast-food industry, such as efficiency and predictability, have influenced society and led to more uniformity. GenAI is different from previous technologies that merely spread information, the researchers say, in that it actively shapes it. Studies suggest that AI tools can flatten language use, creativity and cultural values, and there is evidence that the technology can even influence the decisions we make, the way we act and the opinions we hold. So how big a problem is this? And what can be done to combat it? Although AI homogenization hasn't upended society and might never get to that point, Sourati imagines a worst-case outcome in the long term, in which humans collectively become less adaptable. "That's actually really scary for me." Mind hijacking Emily Wenger, a computer scientist at Duke University in Durham, North Carolina, says the risk of homogenization is an existential one. "If we're all using AI to write our e-mails or whatever, what do we become as a species?" We might fall prey to groupthink, she warns. "We have suffered as a society when we silence or ignore edge voices." Wenger has investigated the creative homogeneity of AI models. She and a colleague tested 22 LLMs and 102 people on three creative tasks, including one on divergent thinking that asked the models and participants to name alternative uses for common objects. The pair found that LLMs produced ideas that were slightly more original -- more semantically different from the question -- than people did. But the LLMs' responses were more similar to each other than the human responses were. Other researchers mapped the narrative features of short stories written by people and five LLMs by analysing the plots, characters, settings and other elements. They found that the AI-generated stories clustered together on the basis of their narrative features, but that those in the human-written ones were more spread out. Working with AI tools can also reduce human creativity. In a 2024 study, people wrote eight-sentence short stories on assigned topics. Some could first ask an LLM for ideas. AI use led to stories that evaluators rated as more novel and enjoyable (except for those by the most talented writers). But it also made stories more similar to each other. In a 2025 study that involved brainstorming tasks, the ideas people submitted were more creative (according to human evaluators) when they used ChatGPT, but again AI use reduced the diversity of the ideas. Recent meta-analyses combining several studies have supported these observations. One published in April found that AI homogenization was strongest for idea generation, especially in complex or constrained tasks. Similar effects are showing up outside laboratory experiments. One study examined more than 400,000 scientific articles in the Web of Science database and found that after ChatGPT's release in late 2022, the number of articles per author increased, but so did their similarity in content and linguistic style. Other research has focused on the use of language. Sourati and his colleagues studied local news articles, arXiv preprints and posts on the social-media platform Reddit and found that after ChatGPT's initial release, linguistic measures of writing style showed a decrease in variation. The researchers reproduced this effect by using LLMs to correct the grammar in human-written texts, which erased many signifiers of personality, moral values and demographics. GenAI can also mute cultural differences. In one study, people in India and the United States were asked to describe their favourite rituals, heroes and symbols, and to write an e-mail in a way that would demonstrate their values. Half wrote using an autocomplete tool that suggested up to nine words each time they paused. This use of AI increased writing similarity between Indians, between Americans and between the two groups, leading Indians to sound more American in their wording. It also led to fewer rich details about, say, the Indian festival of Diwali; instead, descriptions in the writing samples were more generic. The output of AI models can affect the ideas of those who view it. One study before the 2024 US presidential election found that LLMs were biased towards the initial Democratic nominee, Joe Biden, over the Republican candidate Donald Trump, and that chatting with the models about politics made Trump supporters less favourable towards him. LLM content can corrupt the cognition of both consumers and creators. Mor Naaman, an information scientist at the research centre Cornell Tech in New York City, who co-authored the study on the use of autocomplete tools, thinks that genAI creates both cultural and cognitive imperialism. The autocomplete feature, he says, leads to "mind hijacking", altering what users think. In one study that he worked on, participants wrote about social media with the help of an LLM that was secretly prompted to argue that social media was good or bad. The LLM affected not only what users wrote, but also their attitudes in a post-writing survey. In a follow-up study, people's beliefs about issues such as the death penalty skewed towards that of the model's prompt even when they were surveyed weeks later, and when they had been warned about the LLM's bias. In work by other researchers, half the participants had access to ChatGPT for five days to assist with tasks that tested their creativity, and half did not. When all participants completed tasks two months later, none of them had AI access, but those with previous AI access still gave answers that were more similar to each other. The researchers called this a "creative scar". Sourati describes the impact of AI on cognition by invoking George Orwell's 1949 dystopian novel Nineteen Eighty-four. "The way that you talk affects the way that you reason," Sourati says. In other words, if AI use leads everyone to talk the same, they might reason the same, too, he says. Taking the average The homogenization of AI output and, in turn, users' thoughts, have causes that are, respectively, technical and psychological, experts say. When a generative model fails to provide a variety of outputs, it's called mode collapse. This can happen for a few reasons: one is that there is limited diversity in the data used to train the models. The second is that they are rewarded for predicting likely patterns, not divergent ideas. A third is that they are fine-tuned to please human raters, who don't usually prioritize novelty in responses. There's another phenomenon called model collapse. Definitions vary, but broadly it means that if a model is trained on output from itself or another model -- as will increasingly happen now that AI tools are filling the Internet with slop -- after a few generations, the output degrades or loses diversity. Even without model collapse, AI systems can actively bias human culture to become more AI-like and homogeneous. Studies have found that LLMs give a higher rating to CVs and works that they have produced than to those written by other models or humans. That could incentivize us to appease them. Psychologically, Naaman says, people tend to follow an LLM's lead because some look to it as being either representative of others' views or a sage that knows what's best. And when they use an LLM in their work, as in the autocomplete studies, they might feel some ownership of its output. People generally adjust their attitudes to match their own self-perceived behaviour, which can lead them to adopt the perspectives of AI models. AI-produced ideas also have an anchoring effect that fixes our thinking, says Alwin de Rooij, a creativity researcher at Tilburg University in the Netherlands.

AI is making us sound the same -- and killing our personal expression

Researchers found that LLMs push writing toward common stylistic norms, reducing linguistic diversity and blurring signals of personality. More than a third of all internet web pages published since ChatGPT's November 2022 launch were authored by AI, according to the Pew Research Center. And the writing is becoming simpler than ever. That flood of AI-generated language may also be changing the way humans write. A new paper in Nature Human Behavior finds that LLMs can push people's writing toward a narrower set of linguistic norms, reducing stylistic diversity and obscuring some of the identity signals embedded in our words. Three years ago, lead researcher Zhivar Sourati read The Secret Life of Pronouns: What Our Words Say About Us, James Pennebaker's book that explored how people's personal values, backgrounds, and beliefs can lead them to communicate the same ideas using different words. At the same time, LLMs were becoming more popular, and Sourati began to wonder whether the tech might reduce that linguistic diversity. "We were kind of asking, what would actually happen when everyone's text is going through the same models?" Sourati says. "If you have this layer on top of everyone's text, would you change any of these artifacts that we know about people's texts? Would it affect it in any significant way? What are the implications?" So, he set to work to find out. First, Sourati conducted a time-series analysis using approximately 80,000 academic papers, 400,000 news articles, and 300,000 Reddit stories written before and after ChatGPT's introduction. He found that after the adoption of AI writing tools, variation in writing complexity increasingly converged toward common stylistic norms. The trend was consistent across all three datasets. Next, Sourati and his team conducted a more traditional experiment. They had GPT-3.5, Gemini, and Meta's Llama 3 rewrite human-authored texts using a variety of prompts. While the models largely preserved the texts' meaning, they reduced variation in writing complexity 21% to 50%. But Sourati wanted to get more specific about the societal implications: How are LLMs eroding the signals of individual identity in our writing?