Creativity is considered a typically human trait. But AI systems are increasingly able to create works of art, write poems and think up stories. Researchers pitted humans and chatbots against each other in a classic creative thinking task. They were asked to think of new, unusual uses for four everyday objects. The generative artificial intelligences performed better than average human test subjects. However, they had not yet come close to the most creative human ideas.
Systems based on artificial intelligence such as ChatGPT can write specialist articles, obtain information and imitate human communication in a deceptively realistic manner. Generative AI has also proven to be surprisingly powerful in the artistic field. The AI systems compose pieces of music, create paintings and write poems that are often difficult to distinguish from human works. So are they competing with us when it comes to creativity?
People versus chatbots
Since creativity has many facets and can hardly be assessed objectively, the answer to this question is difficult. To compare the creativity of humans and chatbots, Mika Koivisto from the University of Turku in Finland and Simone Grassini from the University of Bergen in Norway chose a classic task that tests so-called divergent thinking – a type of original problem solving that is considered a facet of creativity applies. “We asked 256 human test subjects and three AI chatbots to find unusual and creative uses for four different everyday objects,” explains the team. The items included a pen, a box, a rope and a candle.
Both the human test subjects and the AI systems ChatGPT3.5, ChatGPT4 and Copy.ai were instructed to think of alternative possible uses for each of the items. It was explicitly not important how useful or practical the respective application was. Koivisto and Grassini had six people who had previously been trained for this task evaluate the creativity of the answers. The evaluators did not know that some of the ideas were AI-generated. The researchers also used a computer model to determine how semantically the answers differed from the name of the respective object. Larger differences counted as more creative.
AI beats “average person”
The result: “On average, the AI chatbots performed better than the human participants,” report the researchers. “While the human responses also contained low-quality ideas, the chatbots generally provided more creative responses.” Among the chatbots, ChatGPT4 received the highest ratings for creativity. The responses of its predecessor ChatGPT3.5 and the chatbot Copy.ai, which is based on ChatGPT3, were rated as slightly less creative by evaluators. However, the chatbots hardly differed in terms of the semantic distance to the original object. However, in all areas they outperformed the average human participants.
According to the researchers’ analysis, one of the reasons why the chatbots’ answers were rated as more creative on average was probably because they formulated their ideas in a particularly surprising and unusual way. For example, raters gave ChatGPT4’s suggestion of using a box as an “amusement park for cats” a higher score than the human idea of using it to build a “playhouse for cats.” “However, the best human ideas still matched or exceeded those of the chatbots,” the team reports. “However, AI technology is evolving quickly and the results may look different in six months.”
More creative than us humans?
So will artificial intelligence outperform even the most creative people in the future? This cannot be deduced from the current study. “Creativity is a complex phenomenon, and here we only focused on performance on the most commonly used task to measure divergent thinking,” the researchers write. “Our study highlights the potential of artificial intelligence as a tool to enhance creativity, but also the unique and complex nature of human creativity, which is difficult to fully replicate or surpass using artificial intelligence.”
Source: Mika Koivisto (University of Turku, Finland) et al., Scientific Reports, doi: 10.1038/s41598-023-40858-3