The researchers note that within human groups and societies, cognitive diversity is an important factor for creativity and problem-solving. However, they argue that this diversity is reduced when billions of people use the same limited number of chatbots to perform a wide range of tasks.
A new paper by computer science and psychology researchers argues that the widespread use of AI-based chatbots could lead to a homogenization of writing, speaking, and even thinking styles. If this trend continues unchecked, it could reduce the diversity of human thought and society's ability to deal with complex problems.
The claim appears in an opinion article published March 11, 2026, in the journal Trends in Cognitive Sciences from Cell Press. The researchers call on developers of large linguistic models (LLMs) to incorporate a wider range of languages, perspectives, and thinking styles into their training sets.
According to lead author, computer scientist Zhivar Sourati of the University of Southern California, humans differ in how they express ideas and analyze the world.
“People differ from each other in the way they write, think, and interpret reality,” said Surati. “When these differences are mediated through the same large linguistic models, their unique linguistic style, perspective, and thinking strategies become more uniform—and the result is standardized expressions and thoughts across different users.”
Cognitive diversity as an engine for creativity
The researchers note that within human groups and societies, cognitive diversity is an important factor for creativity and problem-solving. However, they argue that this diversity is reduced when billions of people use the same limited number of chatbots to perform a wide range of tasks.
For example, when people use chatbots to improve their writing, the result can be a loss of personal style. Additionally, users may feel less creative ownership of the final product.
“The concern is not just that large linguistic models influence how people write or speak,” Surati said. “They also subtly redefine what counts as credible speech, a correct point of view, or even good thinking.”
Various studies have found that texts generated by large language models are less diverse than texts written by humans. In addition, the outputs of these models often reflect the language, values, and thinking styles of Western, educated, industrialized, wealthy, and democratic societies.
According to the researchers, this is because these models are trained to identify statistical patterns in their training data – data that often represents a relatively narrow portion of the human experience.
It also influences the way we think.
The research also points to effects beyond language. Studies have shown that after interacting with biased linguistic models, users' opinions may move closer to the positions the model presents.
Additionally, large linguistic models tend to encourage a linear style of thinking, such as “chain-of-thought reasoning,” in which a thought process is presented step-by-step. This emphasis may reduce the use of intuitive or abstract forms of thinking, which are sometimes more effective.
The researchers also note that these models can influence users' expectations, thereby finely tuning the direction of their work.
According to Surati:
“Instead of actively leading the creative process, many users tend to accept the model’s suggestions and choose the option that seems ‘good enough’, rather than developing their own idea. Thus, control gradually shifts from the user to the model.”
The need for diversity in artificial intelligence models
The researchers are calling on AI developers to incorporate a wider range of languages, perspectives, and thinking styles into their models. They say this diversity should reflect the true diversity of human society around the world – and not rely solely on random variations.
“If large linguistic models had more diverse ways of approaching ideas and problems, they would better support the collective intelligence and problem-solving capacity of society,” Surathi said. “We need to diversify the AI models themselves and also change the way we use them, especially given their widespread adoption.”
the article:
"The homogenizing effect of large language models on human expression and thought"
magazine: Trends in Cognitive Sciences
DOI: 10.1016/j.tics.2026.01.003
More of the topic in Hayadan:
