A webinar hosted by AI at Wharton brought together researchers from Boston University, the Technical University of Munich, and Purdue University to explore how generative AI is reshaping artistic and creative production. Yannick Exner's research found that AI-generated visual marketing content can match or even outperform human-made images in perception, social media engagement, and click-through rates. Ankit Sisodia presented methods for using deep machine learning to identify appealing product design attributes and generate new products that resonate with consumers. Dokyun Lee's study of over 4 million artworks showed that text-to-image AI enhanced creative productivity by 25% and increased art's perceived value by 50% over time, positioning AI as a powerful brainstorming tool for artists.
The panelists also debated the long-term implications of AI for human creativity. While they acknowledged AI's vast potential, Exner raised concerns about diminishing returns if people rely too heavily on AI models trained on existing human knowledge, potentially eroding the novel thinking that feeds those models. Lee drew a parallel to the invention of photography, which didn't kill art but transformed it, suggesting AI may have a similar effect on the creative landscape. The discussion highlighted both the promise and the tension inherent in integrating AI deeply into creative workflows.
A new study from Wharton researchers reveals a significant trade-off when using ChatGPT for creative tasks: while generative AI improves the quality of individual ideas, it substantially reduces the diversity of ideas across teams. The research, led by Mack Institute fellow Lennart Meincke and co-authored by professors Gideon Nave and Christian Terwiesch, found that in one experiment, only 6% of AI-generated ideas were unique compared to 100% in the human-only group. Participants using ChatGPT frequently converged on strikingly similar concepts, with several independently naming their toy invention "Build-a-Breeze Castle."
The underlying cause is twofold: the AI model draws from the same statistical distribution when given similar prompts, and users tend to frame their queries in similar ways. In 37 out of 45 comparisons, AI-assisted ideas were significantly less diverse than those produced by other methods. The researchers argue that diversity is essential for breakthrough innovation and warn against mistaking fluency for originality.
To mitigate this homogenization, the study recommends several practical strategies: varying prompts deliberately, using chain-of-thought prompting to break tasks into structured steps, starting brainstorming sessions with human ideas before introducing AI, and employing multiple AI models simultaneously. As Terwiesch noted, "Diversity is often overlooked, but it needs special protection — if you don't solve for it explicitly, you won't get it."
