Abstract: LLM-based agents offer new potential to accelerate science and reshape research work. However, the quality of final intellectual outcomes can vary significantly depending on degrees of human involvement. How can we best use these tools to augment scientific creativity without undermining aspects of contribution and ownership that drive research? This work investigates varying levels of human controls (minimum, medium, and intensive) over an agentic research ideation workflow, revealing how they affect AI creative support and the researcher personal effort they invest. Our mixed-methods study with 54 researchers suggests three findings: 1) the perceived creativity support of AI does not simply increase linearly with greater control; 2) human effort remains consistent across control levels but the nature of work shifts from ideating to verifying; and 3) ownership of the final intellectual work becomes a negotiated outcome between human and AI. Future tool design for AI-driven automated research ideation should empower researchers, allowing them to experience a greater sense of ownership over more powerful ideas, rather than reducing them to operators of a machine that controls the creative process.

Bio: Houjiang Liu is a PhD student studying human-centered computing at the school of information. His recent research focuses on examining how LLM-based agents shape different stages of research activities.