In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models

Researchers replicated Picbreeder, a landmark human-driven collaborative art generation platform, by substituting Vision Language Models for human users to test whether AI agents can engage in open-ended creative discovery. The study reveals qualitative differences between AI-generated outputs and historical human baselines, with findings suggesting that factors like exploratory noise, behavioral diversity, and memory mechanisms significantly influence AI creative capacity.

This research addresses a critical question at the intersection of AI development and creative cognition: can artificial agents generate novel, meaningful outputs without explicit human guidance? Picbreeder represents the gold standard for measuring open-ended creativity, having produced diverse imagery through evolutionary processes guided by human aesthetic judgment. By replacing human curators with Vision Language Models, the researchers created a controlled experiment to assess machine creativity at scale.

The study builds on decades of research into open-ended systems and evolutionary art, extending into the era of multimodal AI. As organizations increasingly deploy AI for creative and scientific tasks, understanding the boundaries of machine-driven discovery becomes commercially and philosophically consequential. The researchers' methodical approach—introducing exploratory noise, behavioral diversity, and memory mechanisms—mirrors cognitive science approaches to human creativity, suggesting that open-endedness may require intentional architectural choices rather than emerging naturally from language models.

For the AI development community, the findings carry implications for designing systems intended for autonomous exploration and discovery. The observation of qualitative differences between human and AI outputs suggests that current VLMs, while powerful, approach creative generation fundamentally differently than humans do. This matters for researchers building scientific discovery systems, creative tools, and autonomous research assistants. The availability of open-source code enables reproduction and extension of these findings across different model architectures and creative domains.

• →Vision Language Models demonstrate qualitatively different creative outputs compared to human-driven Picbreeder baseline, indicating distinct approaches to open-ended generation.
• →Exploratory noise, agent behavioral diversity, and memory of past actions emerge as causal factors influencing AI creative capacity and novelty production.
• →Phylogenetic complexity metrics and visual-semantic novelty measures provide quantifiable frameworks for comparing human versus machine creativity.
• →The research suggests open-endedness may require intentional architectural design choices rather than emerging spontaneously from foundation models.
• →Publicly available code enables broader investigation of AI creativity across different model families and creative domains.