Human–AI collaboration via vibe coding

As AI systems become everyday collaborators, we need principled ways to understand how humans and machines coordinate effectively while preserving human intent, values, and agency. This project studies vibe coding: iterative workflows in which users guide AI systems to generate, refine, and debug code using natural language feedback and rapid interaction.

By turning collaborative coding into a controlled experimental paradigm, we make coordination, communication, and control measurable and computationally precise. Our goal is to identify role structures and interaction patterns that reduce misalignment, improve performance, and support creativity—especially as non-experts increasingly produce high-quality work with AI assistance.

What we do

• Build experimental pipelines that simulate multi-round human–AI collaborative coding workflows.
• Use constrained coding media (e.g., SVG generation and editing) to precisely measure communication, control, and refinement.
• Manipulate role definitions and interaction protocols to test their effects on alignment, performance, and learning.
• Analyze how feedback structure shapes convergence, diversity, and robustness in collaborative outputs.

Why it matters

Human–AI collaboration will define future scientific, creative, and technological work. Understanding how to structure interaction so that AI systems amplify rather than override human cognition is critical for building trustworthy, interpretable, and aligned intelligent systems. This work informs the design of collaborative programming tools, creative AI systems, and hybrid human–machine institutions.

Related publications and links

• placeholder: vibe-coding preprint
• placeholder: experimental framework
• placeholder: dataset
• placeholder: code repository