There has been a creeping sense that, since large language models like ChatGPT have become publicly accessible and more widely used, people have started writing and sounding more like AI-generated content. (Of course, AI is just doing an impression of an amalgam of human-created material, so there’s a clear chicken and egg here.) Now there is evidence to support that sneaking suspicion. In a paper published Wednesday in the journal Trends in Cognitive Sciences, researchers at the University of Southern California warned that the use of LLMs risks flattening human thought and creativity.
The team of researchers analyzed more than 130 studies to better understand how large language models affect cognitive diversity, examining research across a variety of fields from linguistics to computer science. The team found that, despite the fact that AI models pull from a huge database of information, they consistently produce outputs that are less varied than human thought.
That is in part because, while these models may be trained on a seemingly endless supply of human-produced thoughts and ideas, they’re not capable of actually processing all that material in a way that considers the diversity of opinion available. Instead, LLMs tend to favor consistent patterns that they can identify in the training data, which is part of the reason some critiques of the 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 overrepresent dominant languages and ideologies, their outputs often mirror a narrow and skewed slice of human experience,” author and computer scientist Zhivar Sourati of the University of Southern California said in a statement.
Some LLMs even advertise this fact. OpenAI explicitly states that ChatGPT is “skewed towards Western views,” for instance, and xAI has, pretty obviously, tweaked its chatbot Grok to reflect the views of CEO Elon Musk on more than one occasion.
The result of interacting with models that significantly favor certain perspectives, though, is that humans then start to internalize and reflect those perspectives. This can be as simple as a person 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 shift the way people think to be more in line with the information being provided to them by a chatbot. LLMs also use chain-of-thought reasoning, 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 the researchers made was that, while individuals using LLMs to generate ideas often produce more volume (albeit with less creativity), groups of people actually produce fewer ideas when using LLMs compared to when they are simply tasked with collaborating and bouncing ideas off of each other. Basically, using the model locks people into a particular way of thinking and reduces the diversity of perspective that might otherwise come out of discussion and sharing experiences.
It’s been well understood for a while now that diversity of thought and experiences produces better outcomes for groups and organizations. That holds true as it relates to LLMs, which are essentially encouraged to seek consensus thought rather than diversity. Don’t expect that problem to get corrected any time soon, either, considering the Trump administration issued an executive order effectively punishing any company that creates an AI model that promotes diversity.
Read the full article here
