Human–AI Collaboration via Vibe Coding | Cornell Computational Cognition Lab

As AI code generators become everyday tools, a pressing question emerges: what is the human’s actual role? Does human guidance improve outcomes, or can AI systems converge on good solutions alone? This project builds a controlled experimental framework for studying collaborative vibe coding: an iterative, natural-language style of programming in which people guide AI systems toward desired outputs without directly writing code.

We use Scalable Vector Graphics (SVG) as the coding medium because SVGs are easy to render, compare, and validate while still capturing the core challenge of translating high-level visual intent into executable code. The task allows us to manipulate who provides instructions, who selects outputs, and how information flows across iterations, making it possible to compare human-led, AI-led, and hybrid forms of collaboration under controlled conditions.

Vibe coding experimental paradigm with instruction, code generation, selection, and validation — Vibe coding experimental paradigm. (A) An instructor views a reference image and the best SVG rendition from the previous iteration, then provides natural-language instructions to guide code generation. The code generator produces SVG code, which is rendered into an image. (B) In the iterated procedure, a selector chooses whether the current or previous SVG image better matches the reference. The selected SVG is passed forward to the next instructor. (C) Human validation experiments ask participants to rate the similarity between generated SVG images and the reference image.

Across experiments, the project asks how labor should be divided in mixed human-AI systems. The results show that people provide uniquely effective high-level instructions for vibe coding, whereas AI-provided instructions can produce rapid drift or even performance collapse. Hybrid systems perform best when humans lead the process by providing instructions while evaluation is delegated to AI.

Multi-round Human–AI vibe-coding versus AI-only trajectories — Human-led and AI-led vibe coding examples. (A) Example progressions for one reference image, comparing human-led chains (top) with AI-led chains (bottom). (B) Final-iteration examples across reference images. Human-led chains more reliably preserve recognizable structure, while AI-led chains often drift away from the target.

What we do

Build an SVG-based experimental pipeline that makes human-AI collaborative coding measurable and causally controlled.
Compare human-led, AI-led, and hybrid workflows across repeated iterations of instruction, generation, selection, and validation.
Quantify when AI systems drift from human intent and when human guidance stabilizes creative search.
Identify which parts of the workflow, including instruction and evaluation, are best handled by humans, AI systems, or hybrid arrangements.

Why it matters

Vibe coding is rapidly changing who can produce software and how creative technology work is done. Understanding when human involvement is essential, and when AI alone falls short, is critical for designing tools and workflows that amplify human agency rather than displace it. This project provides controlled evidence that human guidance can be a central ingredient in productive human-AI collaboration, especially when open-ended goals must be translated into concrete artifacts.

(Hu et al., 2026)

What we do

Why it matters

Related Publications

2026