AI is homogenizing thought

(CN) — Thanks to AI chatbots, variation in online speaking and writing has visibly decreased, which could have serious consequences for humanity’s collective wisdom in turn.

Many frequent users of the internet may feel it’s easy to detect if something has been written by an artificial intelligence chatbot like ChatGPT or Gemini. Whether being used to write dating app profiles, advertisements or even social media comments, these programs can generate passages that have repetitive sentence structures, neutral tones and an artificial feeling.

But despite perceived increased skills at detection, people still only accurately identify if text is AI half the time. And even as online AI detectors have become more prevalent, AI is everywhere and growing fast — so what does that mean for the more than 6 billion people who are unknowingly exposed to it almost every time they’re online?

Funded by the Air Force Office of Scientific Research, computer scientists and psychologists at the University of Southern California undertook the daunting task of analyzing more than 130 studies relating to large language models and cognitive diversity. The group synthesized evidence from research projects ranging from linguistics and psychology to cognitive science and computer science.

Consistently, these studies revealed that writing generated by AI was less varied than human-generated writing. It regularly lacked the unique language, values and reasoning styles seen in writing produced by modern educated societies.

Published Wednesday in the Cell Press journal Trends in Cognitive Sciences, the researchers’ opinion paper has bleak results. They detail a possible future of heavily standardized thought and language and even make direct comparisons to Orwell’s “Nineteen Eighty-Four.”

Unchecked exposure to large language models in their current state, the researchers warn, risks “flattening the cognitive landscapes that drive collective intelligence and adaptability” and would likely reduce humanity’s collective wisdom and ability to adapt.

The uniqueness of how we speak, think and even interact online inspires thought and innovation in a way no AI chatbot can. This cognitive diversity is a primary source of creativity and problem-solving, says Zhivar Sourati, a computer scientist and the paper’s first author.

“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,” says Sourati.

That means that as more and more people turn to programs like ChatGPT for even just editing a report or important email, style and individuality are lost. Over time, credible speech, good reasoning and perspectives are being redefined as people are regularly exposed to information that’s been processed through large language models.

“Because LLMs are trained to capture and reproduce statistical regularities in their training data … their outputs often mirror a narrow and skewed slice of human experience,” says Sourati. However, content produced by AI is often still palatable on face value, making it difficult for users to differentiate between authentic, human-made content and content designed by a robot.

Groupthink is generated as a result. Although studies may show more ideas can be generated when LLMs are used by an individual, many more ideas — and much more creative ones — are generated when people simply work as a group and combine collective skills and perspectives. Diversity in language, viewpoints and forms of reasoning is essential to our creativity and collective intelligence.

Even those who refuse to use AI firsthand will be impacted indirectly, says Sourati. Individuals will consistently begin to feel that their opinions aren’t good enough and instead align with the group and its one consistent opinion, which in turn shifts agency from the user to the model.

“If a lot of people around me are thinking and speaking in a certain way and I do things differently, I would feel a pressure to align with them because it would seem like a more credible or socially acceptable way of expressing my ideas,” he says.

Reasoning methods are also strongly affected by use of LLMs, as the programs frequently use chain-of-thought reasoning to come to conclusions, one of the simplest forms of reasoning. Consistent exposure to this form of thought will reduce the use of intuitive or abstract reasoning styles, researchers say, which generate creativity and innovation and can be more efficient than linear reasoning.

The paper explains that more real-world diversity, such as real-world global perspectives, should be integrated into the LLM training sets to inhibit these complications. Doing so would also preserve human cognitive diversity and improve skills of chatbots, such as reasoning abilities.

Approaching problems and ideas in a more diverse manner would better support our collective intelligence and problem-solving capabilities, says Sourati.

“We need to diversify the AI models themselves while also adjusting how we interact with them … to protect the cognitive diversity and ideation potential of future generations,” says Sourati.