AI chatbots aren’t just imitating human creativity — they may be flattening it, according to a growing body of research highlighted in a recent Nature feature by science writer Matthew Hutson. Large language models tend to converge on similar, safe outputs — a pattern called “mode collapse” — and when models train on AI-generated text, diversity erodes further, a problem researchers dub “model collapse.”
The sameness spreads to us, too
The homogenization doesn’t stay contained to the machines. “The way that you talk affects the way that you reason,” says Zhivar Sourati, a PhD student at the University of Southern California, who’s noticed it in the research literature: “I read papers and I’m like, ‘I’ve seen this paper before.'”
In a study published in PNAS Nexus, Duke University computer scientist Emily Wenger and a colleague tested 22 LLMs against 102 people on creative tasks and found the models’ answers clustered together far more than humans’ did.
Related on the SLP
“If we’re all using AI to write our e-mails or whatever,” Wenger asks, “what do we become as a species?”
|
Related on the SLP
ChatGPT and other generative AI suppresses brain activity. Is it slowly making us stupid?
|
