When
Where
Léonard Boussioux, Assistant Professor of Information Systems & Operations Management, Foster School of Business, University of Washington
One Tool, One Taste? The Hidden Cost of Agentic Creativity
Abstract: Generative and agentic AI have collapsed the cost of creating: anyone can spin up a polished website, publish an AI-agent "skill," or generate a hundred design concepts in minutes. Yet in our field study of 77 students building websites on the same vibe-coding platform, the outputs converged on a handful of visual archetypes, an aesthetic convergence we call authored ignorance: designers not realizing how much their tool is deciding for them. I extend this into multi-agent creative pipelines, where a controlled biomimicry-design experiment shows that more sophisticated agent architectures raise the judged novelty of individual ideas while shrinking diversity across the population of ideas produced. Widening the lens to the agentic ecosystems these tools are assembling into, a systematic study of 100,000+ published AI-agent "skills" — the instruction files now defining what agents are handed to do, and how — finds the same trade-off encoded into the infrastructure itself: the majority of skills run start to finish with no human step, and what determines full automation is not how risky the work is but how easy its output is to check. Together, these studies point to a hidden cost of AI creativity. I'll close with what this means for how we design, evaluate, and govern agentic tools if we want them to expand the space of what gets built, and keep humans meaningfully present in how it gets built.