GoKawiilZhivar Sourati often gets déjà vu while perusing recent research in his field of computer science. “I read papers and I’m like, ‘I’ve seen this paper before,’” he says. Everything’s looking kind of the same, shorn of individual quirks, says Sourati, who is a PhD student at the University of Southern California in Los Angeles.
Sourati suspects artificial intelligence is to blame. Because people are increasingly relying on large language models (LLMs), their writing is growing more similar, he suggests.
Can AI be truly creative?
Now, Sourati and other researchers are putting their anecdotal observations to the test by exploring how generative AI (genAI) — systems that create text, images and other output — might be homogenizing both culture and cognition. In a March paper, he and his co-authors noted that this phenomenon is similar to the concept of ‘McDonaldization’, invoking how characteristics of the fast-food industry, such as efficiency and predictability, have influenced society and led to more uniformity1.
GenAI is different from previous technologies that merely spread information, the researchers say, in that it actively shapes it. Studies suggest that AI tools can flatten language use, creativity and cultural values, and there is evidence that the technology can even influence the decisions we make, the way we act and the opinions we hold.
So how big a problem is this? And what can be done to combat it? Although AI homogenization hasn’t upended society and might never get to that point, Sourati imagines a worst-case outcome in the long term, in which humans collectively become less adaptable. “That’s actually really scary for me.”
Mind hijacking
Emily Wenger, a computer scientist at Duke University in Durham, North Carolina, says the risk of homogenization is an existential one. “If we’re all using AI to write our e-mails or whatever, what do we become as a species?”
We might fall prey to groupthink, she warns. “We have suffered as a society when we silence or ignore edge voices.”
Wenger has investigated the creative homogeneity of AI models. She and a colleague tested 22 LLMs and 102 people on three creative tasks, including one on divergent thinking that asked the models and participants to name alternative uses for common objects. The pair found2 that LLMs produced ideas that were slightly more original — more semantically different from the question — than people did. But the LLMs’ responses were more similar to each other than the human responses were.
... continue reading