The team of researchers analyzed more than 130 studies to understand how large language models affect cognitive diversity, examining research in a variety of fields from linguistics to computer science. The team found that despite the fact that AI models draw from vast databases of information, they consistently produce outputs that are less diverse than humans would think.
This is because, although these models can be trained on a seemingly endless supply of human-generated thoughts and opinions, they are not actually able to process all of that content in a way that considers the diversity of opinions available. Instead, LLMs favor consistent patterns that they can identify in the training data, which is why some critics of models refer to them as a kind of glorified autocomplete.
“Because LLMs are trained to capture and reproduce statistical regularities in their training data, which often over-represent dominant languages and ideologies, their outputs often reflect a narrow and skewed slice of the human experience,” author and computer scientist Zewar Saurati of the University of Southern California said in a statement.
Some LLMs even advertise this fact. For example, OpenAI clearly states that ChatGPT is “skewed toward Western views”, and XAI has, apparently, made changes to its chatbot Grok to reflect the views of CEO Elon Musk on more than one occasion.
However, the result of interacting with models that significantly support certain viewpoints is that humans then begin to internalize and reflect those viewpoints. This could be as simple as someone using a chatbot to polish their writing and remove some of their stylistic choices, but previous research has shown that interacting with LLMs can actually change the way people think and conform to the information provided to them by the chatbot. LLMs also use chain-of-thinking logic, which reflects a linear form of thinking. They are incapable of more abstract styles of reasoning that may require leaps in logic that are not obvious but can be very effective.
Perhaps one of the most interesting observations made by the researchers was that, while individuals using LLM to generate ideas often generated a greater quantity (albeit with less creativity), groups of people actually generated fewer ideas when using LLM than when they were simply tasked with collaborating and bouncing ideas off each other. Basically, the use of models locks people into a particular way of thinking and reduces the diversity of perspective that might otherwise emerge from discussion and sharing of experiences.
It has been well understood for some time now that diversity of ideas and experiences produces better outcomes for groups and organizations. This is true as it relates to LLMs, which are essentially encouraged to seek consensus rather than diversity. Don’t expect the problem to be fixed any time soon, given that the Trump administration has issued an executive order effectively penalizing any company that creates an AI model that promotes diversity.
<a href