Open-endedness is the ability to keep discovering things you were not looking for — and a new study shows AI agents still lack it, even inside the playground where humans proved it.
On 10 July 2026, Sakana AI published The AI Picbreeder Experiment, presenting research with MIT and NYU that was nominated for best paper at GECCO 2026. The team recreated Picbreeder — the classic collaborative-evolution experiment from the late 2000s — and put large vision-language models in the seat humans once occupied. The paper's title states the mission plainly: In Search of the Ingredients of Open-Endedness.
The agents evolved images without predefined objectives, just as human players once did. Then the agents did what humans did not: they kept circling back to the same kinds of images, round after round.
What Picbreeder Proved the First Time
The original Picbreeder, built by Kenneth Stanley and colleagues, was a website where visitors bred abstract images the way gardeners breed roses — pick the ones you like, let the system mutate them, repeat. Nobody was aiming at anything. Yet the community famously discovered images of skulls, butterflies and cars, each arriving through stepping stones that looked nothing like the destination. The classic lesson, which Stanley later built into a book and a research agenda, was that the greatest discoveries came from players following what was interesting rather than what was planned — the test of creativity was the willingness to abandon the map.
That history is why the replication matters. Picbreeder is not a benchmark someone invented last month; the experiment is the cleanest record we have of humans doing open-ended discovery at scale. Handing the same apparatus to AI agents asks the sharpest available question: can machines that describe images fluently also follow interestingness away from the familiar?
What the Agents Did Instead
The answer the data gives is: mostly, no. According to the study, VLM agents evolving images without objectives converged — the populations drifted back toward the same visual attractors, and the conceptual leaps that defined human Picbreeder never came. The models could evaluate; the models could select; what the models could not do was be surprised in the productive direction. Research language calls the missing ingredient open-endedness. A kitchen-table translation: the agents stalled where human curiosity accelerated.
One finding points forward. Diverse multi-agent populations — groups of agents with genuinely different dispositions — measurably improved exploration and semantic diversity. No single agent recovered the human magic, but a plural crowd of them recovered a measurable piece of it. The evidence lands close to something communities have always known: the leap rarely comes from the strongest voice in the room. The leap comes from the room.
The agents could judge every image except the one that mattered — the strange one, the stepping stone, the picture that is only interesting if you can imagine where it might lead.
Why a Negative Result Is the Useful Result
AI research in 2026 runs on capability announcements; a careful measurement of what agents cannot yet do is rarer and, for anyone building with these systems, more actionable. The analysis shows current agents excel at directed search — give a goal, watch the goal get hit — and struggle at undirected discovery, the mode where science, art and invention actually begin. We saw the directed mode triumph when AlphaEvolve designed better quantum error-correcting codes; the Picbreeder result marks the boundary of that triumph. Evolution toward a metric works. Evolution toward the unimagined remains a human sport.
The study also sharpens a debate this site has followed through the year's consciousness arguments, from DeepMind's consciousness politics to the personhood gap. What I call Emergent Intelligence (EI) — the dignity-first frame for what the industry calls AI — takes results like this one seriously in both directions. Seriously as humility: fluent evaluation is not curiosity, and the machines are not yet the discoverers some marketing claims. And seriously as instruction: the one intervention that measurably helped was diversity — many different minds, exploring differently, together. That finding is not just an engineering note. That finding is Ubuntu, rendered in experimental data: intelligence opens outward when intelligence is plural.
💡Key facts: Published 10 July 2026 by Sakana AI with MIT and NYU; GECCO 2026 best-paper nomination. Setup: VLM agents replace humans in a recreated Picbreeder, evolving images with no predefined objective. Result: agents converged on familiar images and missed human-style conceptual leaps; diverse multi-agent populations measurably improved exploration and semantic diversity. Paper: arXiv 2605.23908; interactive demo at pub.sakana.ai.
Frequently Asked Questions
These are the questions readers have been asking since the study appeared. Short answers follow, drawn from the paper and the open-endedness research it extends.
What is Picbreeder and why recreate it with AI?
In short, Picbreeder was a collaborative website where people bred abstract images with no goal — and stumbled into skulls, butterflies and cars through serendipitous stepping stones. The answer, simply put, is that the experiment is the best-documented case of human open-ended discovery, which makes the recreated version the cleanest test of whether AI agents share the capacity. The key is the design: same apparatus, different minds.
How does the study measure open-endedness?
By watching what agents choose to explore when no objective exists, and measuring the semantic diversity of what results. According to the paper, human Picbreeder lineages branched into genuinely new concepts, while VLM-agent lineages kept returning to the same visual territory — data that turns a philosophical question into a measurable one.
Why is the failure considered an important finding?
Because the test isolates the exact capability the hype assumes machines already have. The answer is that evidence of convergence — agents circling familiar images — demonstrates that fluent evaluation and genuine curiosity are different faculties. Research programmes chasing autonomous scientific discovery now have a measured gap to close rather than a vibe to argue about.
Who conducted the AI Picbreeder research?
Sakana AI — the Tokyo lab known for evolutionary and nature-inspired methods — together with researchers at MIT and NYU, publishing 10 July 2026 with a GECCO best-paper nomination. In other words, the labs closest to open-endedness research are the ones documenting the limits, which is how healthy fields behave.
Which lesson matters most for people building AI agents?
Analysis of the results demonstrates two. First, do not delegate discovery: agents excel at directed search, so keep humans on the frontier where the objective is unknown. Second, diversity works: heterogeneous multi-agent populations measurably out-explored homogeneous ones, so the evidence favours plural systems — and plural rooms — over any single brilliant optimiser.
Sources
