Jivar Sourati often experiences déjà vu when reading new computer science research. “When you read a paper, you think, ‘I’ve seen this paper before,’” says Sourati, a doctoral student at the University of Southern California in Los Angeles. “Everything looks the same, and individual quirks are disappearing.”
Sourati believes artificial intelligence may be contributing to this growing uniformity. As researchers, students, and professionals increasingly rely on large language models (LLMs), their writing may become more similar in tone, structure, and vocabulary.
Can AI really be creative?
Sourati and other researchers are now examining these observations systematically. Their work explores how generative AI—systems that produce text, images, and other forms of content—may be homogenizing culture, language, and human thought. In a March paper, Sourati and his co-authors compared this process with “McDonaldization”: the spread of efficiency, predictability, and standardization throughout society.1
Unlike earlier technologies that primarily distributed information, generative AI actively shapes the information people read and produce. Research suggests that AI tools can flatten language, reduce creative diversity, and influence cultural values. They may also affect the decisions people make, the way they behave, and the opinions they hold.
How serious is AI-driven homogenization? Although it has not permanently transformed society—and may never do so—Sourati worries that the long-term consequence could be a reduced ability to adapt. “That’s actually very scary to me,” he says.
AI and the risk of groupthink
Emily Wenger, a computer scientist at Duke University in Durham, North Carolina, describes the risks of cultural and cognitive homogenization as potentially existential. “What would happen to us as a species if we were all using AI to write things like emails?” she asks.
Wenger warns that widespread dependence on AI could encourage groupthink. “We have suffered as a society by silencing and ignoring marginal voices,” she says.
To study whether AI reduces creative diversity, Wenger and her colleagues tested 22 LLMs and 102 people on three creative tasks. One task measured divergent thinking by asking both the models and human participants to suggest alternative uses for a common object.2 The LLMs produced ideas that were slightly more original on average than the human responses. However, the AI-generated answers were considerably more similar to one another than the answers provided by people.
Another research team analyzed the narrative features of short stories written by people and five LLMs, including plot, characters, settings, and other elements.3 The AI-generated stories formed tight clusters with similar characteristics, whereas the human-written stories were more widely dispersed.
Warnings of AI doom are growing louder. Are they realistic?
Using AI can also affect the creativity of human users. In a 2024 study, participants wrote eight-sentence stories on an assigned topic. Some first asked an LLM for ideas. Although AI-assisted stories were rated as more novel and entertaining—except among the most talented writers—they were also more similar to one another.4
A 2025 brainstorming study produced a similar result. Participants generated ideas that human evaluators considered more creative when they used ChatGPT, but the overall range of ideas became less diverse.5 A meta-analysis published in April, which combined findings from multiple studies, also concluded that AI homogenization is particularly strong during idea generation, especially for complex or highly constrained tasks.6
The effect is visible beyond controlled experiments. One study examined more than 400,000 scientific papers in the Web of Science database. After ChatGPT was released in late 2022, the number of papers published per author increased, but the content and language style of those papers also became more similar.7
Other studies have focused on changes in language. Sourati and his colleagues examined local news articles, arXiv preprints, and posts on Reddit. They found that variation in several measures of writing style declined after ChatGPT’s release. When the researchers recreated the effect by using an LLM to correct grammar in human-written text, many signals of personality, moral values, and demographic identity disappeared.8
Generative AI may also narrow cultural differences. In one study, participants in India and the United States wrote emails about their favorite rituals, heroes, and symbols in a way that reflected their values.9 Half of the participants used an autocomplete tool that suggested up to nine words whenever they paused.
AI assistance made the writing of Indian and American participants more similar. Indian participants’ language also became more aligned with American usage. For example, descriptions of the Indian festival Diwali contained fewer specific cultural details and became more general after participants used the autocomplete tool.
The concept of “McDonaldization,” in which society becomes more standardized, may also describe the growing influence of AI.Credit: Poly Fei/SOPA Images/LightRocket/Getty
AI-generated content can also influence the people who encounter it. A study conducted before the 2024 US presidential election found that LLMs were more favorable toward Democratic candidate Joe Biden than Republican candidate Donald Trump. Conversations with an AI model also made Trump supporters less favorable toward him.10
LLM-generated language may affect both consumers and creators. Mol Nerman, an information scientist at Cornell Tech in New York City and co-author of the autocomplete study, argues that generative AI can produce cultural and cognitive imperialism.9 Autocomplete suggestions, he says, can cause “thought hijacking” by subtly steering what users write and believe.
In one study, participants wrote about social media with help from an LLM that had been secretly prompted to argue that social media was either beneficial or harmful.11 The model influenced not only the participants’ writing but also their attitudes in surveys completed afterward. In follow-up surveys conducted weeks later, participants’ views on issues such as the death penalty remained influenced by the model’s prompts—even when participants had been warned about potential AI bias.12
Another study gave half of its participants access to ChatGPT for five days while they completed creative tasks. The rest received no AI assistance. Two months later, all participants completed another task without access to AI. Those who had previously used ChatGPT continued to produce answers that were more similar to one another. The researchers described this lasting effect as a “creative scar.”7
Sourati compares the potential impact of AI on human thought with George Orwell’s 1949 dystopian novel 1984. “The way you speak affects the way you reason,” he says. If AI encourages everyone to express ideas in the same way, people may also begin to draw the same conclusions.
Why AI outputs become more similar
Experts point to both technical and psychological explanations for the growing similarity of AI-generated content and human thinking. When a generative model repeatedly produces similar answers instead of a broad range of possibilities, researchers refer to this problem as modal collapse.
One possible cause is limited diversity in the data used to train AI models. Another is that models are optimized to predict likely patterns rather than generate genuinely unconventional ideas. Reinforcement learning and human feedback may further encourage uniformity because human evaluators often reward answers that are clear, familiar, and useful rather than surprising or original.
A related concern is model collapse. Although definitions vary, the term generally describes the loss of quality and diversity that can occur when an AI model is trained on content generated by itself or by other models. As AI-generated material fills the internet, future systems may increasingly learn from synthetic data instead of diverse human-created sources.
Scientists have invented fake diseases. AI told people it was real
Even when an AI model does not technically collapse, it can still push human culture toward greater uniformity. Studies have found that people rate resumes and other work written by LLMs more favorably than content produced by humans or other AI systems.13 This could create an incentive for people to write in ways that please AI evaluation systems.
There are psychological reasons people may follow an LLM’s recommendations. Users often view AI systems as knowledgeable, objective, or representative of collective opinion. When people use an LLM to complete a task, they may also feel a sense of ownership over the resulting content, making them more likely to accept its assumptions.
Nerman says people can gradually adopt the perspectives embedded in AI-generated suggestions. Creativity researcher Alwin de Rooy of Tilburg University in the Netherlands adds that AI-generated ideas can have an anchoring effect, shaping the direction of human thinking before alternative possibilities are considered.
Source: www.nature.com