level: research
scientists are building ai assistants to automate creative and scientific work. a key feature of human creativity is open-endedness, the ability to produce endless novel and meaningful outputs. to see if ai can do this, researchers turned to picbreeder, a classic system where users evolved images by selecting and breeding small neural network outputs. they replaced human users with frontier vision-language models (vlms) to see if the machines could drive similar open-ended discovery.
the team built a replica of picbreeder and let vlms guide the evolution of images. they compared the machine-generated image libraries to the historical human-created ones. the results showed clear qualitative differences. the vlm-driven process did not match the diversity and novelty seen in human-driven evolution. the researchers attempted to measure these differences using phylogenetic metrics, which track the evolutionary relationships among images.
this work highlights a gap between human and machine creativity in open-ended tasks. while vlms can generate images, they struggle to sustain the kind of unguided, divergent exploration that humans naturally do. the findings suggest that current ai lacks the intrinsic drive or aesthetic judgment needed for open-ended creative search. understanding this limitation is important for designing better ai assistants for scientific and artistic discovery.
why it matters: understanding ai's limits in open-ended creativity helps improve tools for automated discovery in science and art.