Human creativity is a key to our success – it is what made us what we are today. But what happens if we rely more and more on artificial intelligence for creative ideas and innovations? Does this make us more creative – or does this weaken our species’ ability to innovate? An experiment with 22 different AI models and a good 100 human test subjects now shows: Taken individually, the AI systems deliver creative solutions that can outperform human ones. Overall, however, the AI solutions are far more similar and therefore less versatile than those of humans. This means that creative range and originality at the level of entire groups or even humanity could suffer in the long term if we rely too heavily on AI.
Creativity is considered the key driver of human development and culture. Because it is only this that enables us to create new things and find innovative solutions. Through creative thinking we recognize new connections, think “out of the box” and solve problems with the help of creative brainwaves. These skills help us overcome difficulties, improvise, and adapt to changing conditions. Especially today, creative ideas in research and technology are more important than ever, because without them the big questions of our time probably cannot be solved. In the meantime, however, the conditions have changed fundamentally: new ideas no longer arise in secret or through purely human thought processes: advances in artificial intelligence and especially in large language models (LLM) such as ChatGPT, Gemini or Claude now enable us to include AI systems in our creative processes.
22 AI models and 100 people in the test
But what consequences does the use of AI have for our creativity? “Many studies have already shown that large AI models like ChatGPT actually outperform human test subjects in standardized tests of creativity,” report Emily Wenger from Duke University in the USA and Yoed Kenett from the Israel Institute of Technology in Haifa. This applies both to tests of divergent creativity, which involve, for example, unusual uses for objects, as well as to creative association – solutions by developing connections. But AI models also seem to be helpful in artistic creativity. One study showed that authors developed more creative stories when they used GPT-4. However, it also revealed: “The stories written by different people using AI were more similar to each other than the texts that only emerged from the human mind,” said the researchers. This could mean that the use of artificial intelligence makes us more creative as individuals, but not the human population as a whole.
However, it remained unclear whether this restriction also applies when different AI models are used. “One might assume that these homogenization effects can be avoided by using different AIs,” say Wenger and Kenett. To investigate this, they had 22 different LLMs and 102 test subjects complete three established creativity tests. The Alternative Uses Test (AUT) is about listing as many different uses as possible for a common object – for example a book, fork, table or trousers. In the Divergent Association Task (DAT), test subjects are asked to name ten words that differ maximally from each other in every respect. Both tests are used to assess divergent creativity. The third test, Forward Flow (FF), is intended to test associative, convergent creativity. The aim is to write down as long a chain of associatively linked words as possible.
More creative individually, but homogeneous as a group
The tests showed: In the two tests of divergent creativity, the large language models performed slightly better than the human test subjects, and slightly worse in the associative test. “Overall, humans and AI models show approximately the same level of individual originality,” report Wenger and Kenett. However, that changed when the researchers examined their core question: How creative are the solutions when you compare the results with each other? This showed that for both forms of creativity: “The LLM answers are much more similar to each other than those of the human participants,” reports the team. “Even when we changed the prompts to ask the AI model to be more creative, the human responses remained more variable.” This also applied when different model types or AI systems from different manufacturers were compared.
The reason for the homogeneous outputs of the AI models is obvious: “Commercial LLMs were all trained with the same data – the entire Internet,” says Wenger. “It therefore only seems logical to me that this also limits diversity.” The big question, however, is whether this can be changed in the future. Because true creativity is based on parameters such as authenticity, context, intentionality and individuality. “These are dimensions that cannot be implemented in the current LLM landscape,” say the researchers. In their opinion, the use of artificial intelligence for creative solutions could therefore have negative consequences for our ability to innovate as a species. “Research over the last few decades has repeatedly shown how much human creativity depends on the diversity of solutions,” says Kenett.
Source: Emily Wenger (Duke University, Durham, USA) and Yoed Kenett (Israel Institute of Technology, Haifa), PNAS Nexus, doi: 10.1093/pnasnexus/pgag042