AI programs standardize the way we speak, write and think: specialists

Madrid. Artificial intelligence (AI) chatbots are standardizing the way people speak, write and think, according to work from the University of Southern California.

If this homogenization continues unchecked, it risks reducing the collective wisdom and adaptive capacity of humanity, argue computer scientists and psychologists in an opinion article published in the journal Trends in Cognitive Sciences, de Cell Press.

They claim that AI developers should incorporate greater real-world diversity into the training sets of broad language models, not only to preserve human cognitive diversity but also to improve the reasoning ability of chatbots.

“People differ in the way they write, reason, and see the world,” contextualizes the first author of the article and computer scientist Zhivar Sourati, from the University of Southern California.

“When these differences are mediated by the same broad language models, their distinctive linguistic style, perspective, and reasoning strategies become homogenized, producing standardized expressions and thoughts for all users.”

Within human groups and societies, cognitive diversity drives creativity and problem solving, researchers say.

However, cognitive diversity is declining around the world, as billions of people use the same AI chatbots for an increasing number of tasks. When individuals use chatbots to polish their writing, for example, their stylistic individuality is lost and people feel less responsible for their creative output.

“The concern is not just that broad language models shape the way people write or speak, but that they subtly redefine what counts as credible speech, correct perspective, or even good reasoning,” Sourati warns.

biased experience

The team points to multiple analyzes showing that the results of LLM studies are less varied than human-generated writings and that they tend to reflect the language, values ​​and reasoning styles of Western, educated, industrialized, wealthy and democratic societies.

“Because broad language models are trained to capture and reproduce statistical regularities in their training data – which often overrepresent dominant languages ​​and ideologies – their results often reflect a narrow and biased slice of the human experience,” says Sourati.

Although studies show that individuals often generate more ideas with more detail when using broad language models, human groups produce fewer and less creative ideas when they rely on them than when they simply combine their collective powers, the researchers say.

“Even if people are not direct users of these models, they will affect them indirectly,” Sourati emphasizes. “If a lot of people around me think and speak a certain way and I do things differently, I would feel pressure to align myself with them because it would seem like a more credible or socially acceptable way to express my ideas.”

Beyond language, studies have shown that after interacting with broad biased language models, people’s opinions become more similar to the model they used.

“These schemes also favor linear modes of reasoning, such as chain reflection, which requires models to show reasoning step by step. This emphasis reduces the use of intuitive or abstract reasoning styles, which are sometimes more efficient than linear,” the researchers say. They also point out that broad language patterns can alter people’s expectations, which can subtly change the direction of an individual’s work.

“Instead of actively directing the generation, users often get carried away by the model’s suggested continuations and select options that seem ‘good enough’, rather than creating their own, gradually shifting the initiative from the user to the model,” explains Sourati.

The researchers say that AI developers should intentionally incorporate diversity of language, perspectives and reasoning into their models. They emphasize that this variety should be based on that which exists in humans on a global scale, rather than introducing random variations.

“If broad language models had more diverse ways of approaching ideas and problems, they would better support the collective intelligence and problem-solving capabilities of our societies – adds the specialist -. We need to diversify the artificial intelligence models themselves and at the same time adjust our interaction with them, especially given their widespread use in various tasks and contexts to protect the cognitive diversity and ideation potential of future generations.”