Sakana AI Replaced Human Artists With AI and Exposed a Creativity Gap
Sakana AI replaced human users in the classic Picbreeder image-evolution experiment with VLM agents, revealing a fundamental gap in AI open-ended creativity.

- Sakana AI, MIT, and NYU replaced human users in the classic Picbreeder image-evolution experiment with VLM agents, testing AI open-ended creativity.
- 10 VLM agents ran in parallel, evolving images encoded as small neural networks (CPPNs) with no target or reward signal; Gemini 2.5 Pro performed best.
- Humans still lead: the human archive scores 0.681 on Visual Coverage vs. 0.665 for the best AI run, and produces more semantically diverse lineages.
- Injecting 1,000 distinct agent personalities is the most effective intervention, closing the Semantic Coverage gap to match humans (0.734 vs. 0.730).
- VLMs get trapped in semantic attractors, making small incremental refinements rather than the large conceptual leaps humans make from lucky accidents.
- Code, data, and an interactive demo are all publicly released; paper nominated for best paper at GECCO 2026.
What happens when you take one of the most fascinating human creativity experiments ever run on the internet, and replace every human with a frontier AI model? That is exactly what a team from Sakana AI, MIT, and NYU set out to answer. Their paper, nominated for a best paper award at GECCO 2026, revisits Picbreeder -- a now-defunct website where thousands of users collaboratively evolved abstract images into recognizable art, with no goal, no target, and no instructions. The results are a fascinating and humbling look at what AI still cannot do.
The Original Experiment: Fishing for Images
Picbreeder was built around a deceptively simple idea. It put an interactive evolutionary interface in the hands of the public so that every lineage ever produced by anyone could be explored and continued by anyone else, making users participants in a massive online collaboration to explore the space of possible images. Users did not draw or design anything. Instead, they selected images they found interesting from a grid of candidates, and those selections drove an evolutionary algorithm forward.
The images themselves are not pixels in the traditional sense. Each one is encoded as a Compositional Pattern-Producing Network (CPPN) -- a small neural network that takes pixel coordinates as input and outputs a color. Think of it as a recipe for an image rather than the image itself. Through the process of branching from other images, and through continually increasing image complexity made possible by the underlying NEAT algorithm, evolved images proliferate unlike in any other interactive evolutionary computation system. NEAT (NeuroEvolution of Augmenting Topologies) is the key: it evolves not just the weights of the network, but its structure -- adding new nodes and connections over time, allowing images to grow in complexity across generations.
The magic of Picbreeder was that nobody planned the outcomes. Faces, animals, vehicles, and skulls emerged gradually across many sessions and many different users. Over seven years of operation, the site accumulated almost 10,000 evolved images and over 1,000 users. The best creations were not conceived -- they were stumbled upon.

Swapping Humans for AI Agents
The Sakana team automated the entire Picbreeder loop using vision-language model (VLM) agents. The setup runs 10 agents in parallel, all sharing a growing archive of published images. Each agent session works like this:
- The agent is shown a sample of 100 images from the shared archive, split into 5 subsets.
- It picks one image to "branch" from, starting an evolutionary lineage.
- It breeds and mutates candidates across up to 20 generations, selecting its favorite at each step.
- It publishes its best result back to the shared archive.
- Separately, a "critic" agent periodically rates archive entries on a 1-5 scale, influencing which images get sampled in future sessions.
There is no target image, no reward signal, and no explicit definition of what counts as progress. The archive grows unboundedly. Gemini 2.5 Pro came out on top of the tested models, ahead of Gemini 3 Pro Preview and local Qwen3-VL variants. The team also tested three key interventions -- noise injection, memory length, and agent personality diversity -- to see what levers could push the AI closer to human-level exploration.
Three Knobs That Matter
The researchers identified three interventions that meaningfully change how the agents explore:
- Noise (epsilon): At each selection step, with probability epsilon, the agent picks a random image instead of its preferred one. This acts like a second-order mutation rate, forcing the system out of local attractors. Higher noise improves visual coverage but risks filling the archive with incoherent stepping stones.
- Memory (context length): How many past generations the agent can see. Zero memory causes agents to repeat the same image endlessly. Even one generation of history breaks this loop. But too much memory (10-20 turns) overloads the context and degrades output quality, often producing abstract or repetitive forms.
- Agent personalities (NA): A single personality trait is sampled from a pool and prepended to each agent's system prompt. Scaling from 0 to 1,000 distinct personalities produces the largest gains of any intervention.
The Numbers Tell a Clear Story
The team measured archive quality along three axes: Visual Coverage (how spread out the images look in embedding space), Semantic Coverage (how diverse the captions generated for them are), and Semantic Recall (how many of 1,824 real-world object categories from the THINGS dataset the archive can approximate). The human archive acts as the gold standard.
Measured by Semantic Recall over 1,824 THINGS classes, the best VLM configuration landed at 0.087, marginally trailing the historical human baseline of 0.089. That sounds close, but the gap in Visual Coverage is more telling: humans score 0.681 while the best VLM run with 1,000 agent personalities reaches only 0.665. The diversity of what AI agents actually produce is meaningfully narrower.The personality intervention is the standout result. Scaling from no personalities to 1,000 distinct ones pushes Semantic Coverage from 0.696 to 0.734 -- matching the human score of 0.730. It also dramatically improves tree balance (from 0.235 to 0.476), meaning agents stop obsessively branching from the same few popular parents and explore more of the evolutionary tree.

The Real Gap: Serendipity, Not Skill
The most revealing finding is not in the metrics -- it is in the lineage comparisons. When the team traced how both humans and AI agents arrived at a profile-view car image, the difference was stark. The VLM path went: Chrome Bird, Driver's Seat, Dashboard View, Vintage Chrome, Chrome Insignia, Vintage Hood Ornament, Chrome Roadster, Chrome Speedster, Chrome Streamliner. Every step stayed within the semantic neighborhood of "chrome vehicle." The human path wandered through eyes and alien faces before suddenly landing on an automobile.
Ask any LLM to pick a number between 1 and 10 and you will get a 7 followed by a 3. This is because all LLMs tend to converge on a narrow set of predictable answers, even for open-ended queries. The Picbreeder experiment makes this failure mode viscerally concrete: VLM agents notice interesting patterns, but then get trapped in them. They refine rather than abandon. They optimize rather than wander.
There is also a subtler finding about representational quality. When the team perturbed the weights of a skull CPPN evolved by VLM agents, the image changed smoothly -- less fractured than a skull produced by direct gradient descent optimization, but less cleanly structured than one evolved by humans. Human-evolved skulls have weights that correspond to meaningful semantic controls (mouth opening, eye winking). The AI-evolved skull is somewhere in between: more robust than pure optimization, but still lacking the clean disentanglement that human-guided open-ended search produces. This hints that open-ended evolution -- even AI-driven -- may produce better internal representations than direct optimization, a potentially important signal for future training approaches.
What This Is Actually Useful For
This is not a paper about making better images. It is a diagnostic tool for a much larger question: what ingredients does an AI system need to explore open-endedly, the way humans do in science, art, and mathematics? The practical implications branch in several directions:
- Multi-agent diversity as a creativity lever: The personality injection result is directly applicable. If you are building a multi-agent system for brainstorming, code exploration, or research ideation, seeding agents with diverse behavioral priors -- even simple ones -- substantially increases the diversity of outputs.
- Memory tuning for agentic loops: The sweet spot of one to two generations of history is a concrete finding for anyone designing iterative agent pipelines. Too little context causes repetition; too much causes degradation.
- Open-ended training signals: The representational robustness finding suggests that open-ended evolutionary search could be used as a training signal for better-structured internal representations, without explicitly optimizing for disentanglement.
What Remains Out of Reach
Open-endedness refers to a system's ability to "continuously generate artifacts that are both novel and learnable to an observer." By that definition, the AI agents partially qualify -- but only with significant scaffolding. Without personality diversity, noise injection, and carefully tuned memory, they collapse into repetitive loops. Humans do none of this by design; they just get bored and try something different.
The core limitation is what the authors call the "striver pathology": VLMs are trained to be goal-following, and that training bleeds into supposedly open-ended exploration. Even when agents appear to make a creative leap, the jump is often predictable from an implicit theme. The serendipity that humans bring -- recognizing when a happy accident is worth pursuing, then making a larger conceptual leap -- is not something that emerges from current VLMs, even frontier ones.
The full code and dataset are publicly available, and the team has also released an interactive homage to the original Picbreeder site populated with AI-generated archives. The paper is on arXiv. Whether the missing ingredient is a new training objective, a different architecture, or something we have not yet conceived -- that question is now at least more precisely defined.