GabyAI is developing incredibly rapidly, already outpacing humans in strategic games, web design, and data processing. Now it’s starting to beat us on creative thinking too.
AI generates creative ideas better than humans
AI chatbots outperformed humans when asked to come up with alternative uses for everyday objects. They used a skill known as divergent thinking, which generates creative ideas by exploring many possible solutions.
A Norwegian study involved 256 volunteers and three AI systems — GPT-3, GPT-4, and Copy.ai — which were asked to provide ideas for reusing a rope, a box, a pencil, and a candle. When evaluated using an alternative uses task, which asks people to list as many ways as possible to use a simple object, the chatbots on average performed better than the human participants.
However, the study also found that the top human ideas matched or beat the best AI ideas. In the Daily Mail, Simone Grassini of the University of Stavanger (Norway) highlighted the striking ability shown by the chatbots.
The study concludes that AI outperforms most people on this type of creative task. But the experiment measured divergent thinking, not creativity in every form, and Grassini notes that the most creative people still outperform AI for now.
How was the research conducted?
Each AI was tested 11 times across four prompts in different sessions, while the human participants produced 1,024 observations in a single session. Humans were asked to focus on the quality rather than the quantity of responses and to come up with original, creative ways to use each object.
The goal was to produce ideas that could be clever, unusual, humorous, innovative, or otherwise creative. The ideas did not need to be practical or realistic; they could be nonsensical or strange, so long as they showed a nonconventional use.
Participants could generate as many ideas as they wanted, but assessors prioritized the quality of creativity. A few truly insightful ideas were better than many unoriginal ones. Participants had 30 seconds to respond for each item.
The instructions for the AIs matched the human instructions, with one practical exception: unconstrained GPT-3 tended to output 10 ideas, GPT-4 produced 7–8 ideas, and Copy.ai varied more. To align the number of AI outputs with the human distribution, the researchers limited the number of AI responses based on the human idea counts.
Creative ideas from AI are of higher quality than those from humans
The team found that, on average, chatbot responses scored higher for semantic distance and creativity than human responses. The single best human response beat the single best response from each chatbot in seven out of eight evaluation categories, however human responses included a larger share of low-quality ideas.
Grassini says some chatbot responses demonstrated a very high level of creativity. She expected the chatbots to perform well, but not this well.
While this research highlights AI’s potential as a tool to boost creativity, it also underscores the unique and complex nature of human creativity, which may be hard to fully replicate with current AI systems.
Grassini says it remains unclear whether these AI abilities will be deployed to replace creative human work. She believes, though, that AI can help people enhance their creative skills.
