Generative artificial intelligence (AI) is revolutionizing many aspects of our lives, from customer support to artistic creation. A new study published in Science Advances provides insight into how these AI systems, specifically large language models, impact human creativity in writing. The findings suggest that AI can enhance the perceived creativity and quality of short stories, particularly for less inherently creative writers, but it also raises concerns about the potential homogenization of creative outputs.
Creativity is a cornerstone of human expression and innovation, yet the advent of generative AI technologies has begun to challenge traditional views on the uniqueness of human-created content. In their new study, Anil Doshi (an assistant professor at UCL School of Management) and Oliver Hauser (a professor and deputy director of the Institute for Data Science and Artificial Intelligence at University of Exeter) aimed to investigate how generative AI affects individuals’ ability to produce creative written content, specifically focusing on short fiction.
“We were both excited by the potential of generative AI,” the researchers told PsyPost. “We both thought there would be an opportunity to work in an area of common interest. Why we focused on the question of creativity: because generative AI is such a new and potentially transformative technology, we wanted to focus on a core characteristic of being human—that is, our ability to be creative and express new ideas and output.”
The researchers recruited 500 participants from the Prolific platform, an online research participant pool. They ensured a reliable sample by including only participants with a high approval rating and based in the United Kingdom. After accounting for dropouts and exclusions, 293 participants completed the study.
Participants were randomly assigned one of three writing topics: an adventure on the open seas, an adventure in the jungle, or an adventure on a different planet. They were instructed to write an eight-sentence story suitable for a teenage and young adult audience. The participants were further divided into three groups based on the availability of AI assistance:
- Human-only group: This group wrote their stories without any AI input.
- Human with one AI idea group: Participants could request a single, three-sentence story idea from OpenAI’s GPT-4 model.
- Human with five AI ideas group: Participants could request up to five story ideas from the same AI model.
After completing their stories, participants rated their own work on various stylistic attributes, including creativity and enjoyability. The stories were then evaluated by a separate group of 600 individuals from the same online platform. These evaluators assessed the creativity, quality, and originality of the stories without knowing whether the stories were written with AI assistance.
The researchers found that stories written with access to AI-generated ideas were rated higher in creativity, quality, and enjoyability compared to those written without AI assistance. This enhancement was particularly notable among participants with lower inherent creativity. For these less creative writers, having access to multiple AI ideas resulted in substantial improvements in both the novelty of their stories. These improvements brought their work to a level comparable to that of more inherently creative participants.
“We find that getting ideas from generative AI improves the creativity of a story,” Doshi and Hauser told PsyPost. “What surprised us was that almost all of the increase in creativity was experienced by the least creative writers in our sample. Not only that, but getting multiple AI ideas put the assessed creativity of their stories on par with those who are the most creative in our sample. We saw a clear ‘level the playing field’ effect of getting AI ideas on the creativity of the story.”
A downside of using AI-generated ideas, however, was the increased similarity among the stories. The researchers found that stories from the AI-assisted groups were more alike both to each other and to the AI-generated ideas. This raises concerns about the potential homogenization of creative outputs if AI tools become widely used. The increased similarity suggests that while AI can enhance individual creativity, it might do so at the expense of collective diversity and novelty in creative works.
Another interesting finding was the discrepancy between participants’ self-assessments and the external evaluations of their stories. Participants who used AI assistance did not rate their own stories as more creative or enjoyable compared to those who did not use AI. However, external evaluators consistently rated the AI-assisted stories higher. This suggests that individuals might not fully recognize the enhancements provided by AI to their creative outputs.
“Generative AI tools, like ChatGPT, improve the average creativity of a writer’s story, but collectively, stories that had AI ideas looked more like one another than those that did not receive AI assistance,” Doshi and Hauser said. “So there are potentially significant implications — both positive and negative—for individuals and society as a whole.”
The researchers added that the findings point to a social dilemma: “Individual stories are evaluated as being more creative, so people looking to improve their writing might turn to AI. But, if we all do so, then the collective novelty of ideas decreases, which may not be desirable from society’s viewpoint.”
The study highlights both the potential benefits and risks of AI-assisted creativity. But as with all research, there are some caveats to note.
“Our study included a specific use of AI in order for us to better control the experiment,” the researchers explained. “We controlled the prompt and we did not allow for participants to interact with the AI. We did so because we did not want to create a situation where, say, better writers can provide better prompts to elicit better ideas from AI and they write better stories. That would ‘break’ our goal of identifying a causal effect of AI ideas on creativity. So, there is opportunity to build on our work and understand how different prompts and interactions play a role in the creative process.”
“We are developing a research agenda around generative AI to understand how it might be use in a broad array of economic activities,” Doshi and Hauser added. “For example, we are thinking working on a project to look at how AI assists with creation of new ideas in different settings, such as the development of a company’s strategy. We are also looking at how different types of people might respond differently to generative AI. Overall, our goal is to provide research that organizational and societal leaders can use when considering their own AI policies and strategies.”
The study, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” was published July 12, 2024.
