Analysis: AI models less creative than many people, even as a group
Those who value originality and creativity should continue to rely on humans, not AI chatbots. At least, that's what a new study suggests.
(Image: Alones/Shutterstock.com)
Creative outputs from different AI text generators differ less from each other than content from humans given the same tasks. This was discovered by a research group at Duke University in Durham, North Carolina. So if people wonder whether different AI chatbots can lead them in different creative directions with the same prompts, the answer is “in principle no”: “As a group, LLMs [large language models] are less creative than humans,” summarizes study leader Emily Wenger, the study's findings. This could have negative long-term consequences for human creativity.
Humans are more creative in the aggregate
For the research work, the group pitted 22 LLMs against more than 100 humans in three standardized tests to determine creativity. The first involved listing as many different uses for an object as possible, the university explains. For example, it should be listed that a book can also be used as a doorstop, fly swatter, or fire starter. In the second, participants and the AI models were to list 10 different words that should differ from each other as much as possible. Finally, everyone was to write down what came to mind for a word. This had to be repeated until a series of 20 words was created.
Together, these exercises are intended to measure “divergent and dissociative thinking abilities” that promote creativity, it continues, with a clear result: While individual LLMs can outperform individual humans in terms of creativity, the entirety of AI-generated responses was much more similar to each other than those of humans. In the Alternative Uses Test and the Association Test, individual LLMs even slightly outperformed individual humans, while humans performed better in the Association Chain Test (Forward Flow). If the prompts were adjusted to elicit more creative answers from the technology, this only slightly changed the variance. Even then, humans won. Wenger adds that she suspected this; after all, all AI models are trained on roughly the same data material – the entirety of the internet – which should level their outputs.
Videos by heise
“This study has far-reaching implications as people increasingly integrate LLMs into their daily lives,” says Wenger. Over-reliance on these tools will lead to “the world's language becoming increasingly confined to the same vocabulary and grammar, leading to texts becoming more and more similar.” Anyone who wants to develop an original concept or product should, according to the analysis, “bring together a diverse group of people to brainstorm rather than relying on AI.” The complete research paper has now been published in the journal PNAS Nexus.
(mho)
