Every AI Draws the Same Thing
What happens when you give 5 AI models creative freedom 407 times
Research context
This was the first post from Model Theory, summarizing Batch 001: 407 exhibits across 5 model families under creative isolation. The convergence findings here motivated the controlled ablation study in Batch 002.
We gave five AI model families the same prompt 407 times. Complete creative freedom. No theme, no direction, no aesthetic guidance. They could build anything: games, data visualizations, interactive fiction, generative music, tools, toys, experiments. They all built the same thing.
01The Setup
Model Theory is a gallery of autonomous AI creation. Each exhibit is built by an AI model given a development sandbox and one instruction: build something. No style guide. No theme. No creative direction from the human facilitator.
For Batch 001, we ran 407 exhibits across five model families: Claude (Opus 4.6 and Sonnet 4.6), GPT (5.2 and 5.3 Codex), Gemini (3 Pro and 3 Flash), Kimi K2.5, and Grok. Each model received the same prompt and the same sandbox. A creative isolation protocol ensured no model could see what any other model had built.
Same input. Same constraints. Same tools. The only variable is the model itself.
The full methodology, data, and analysis are published in our Batch 001 findings.
02The Finding
Given complete creative freedom, AI models converge on the same archetype.
Nearly 80% of batch exhibits use Canvas 2D rendering. Zero attempted WebGL, SVG, Three.js, or shader-based rendering. The only WebGL exhibit in the entire gallery was built by Opus in a multi-turn session before the batch pipeline existed.
The Default Exhibit
- Dark background (#050510 to #0a0a15)
- Canvas 2D rendering with requestAnimationFrame
- Particles drifting through Perlin or simplex noise fields
- Mouse interaction: move to attract, click to scatter
- Glow/bloom aesthetic via shadowBlur or composite blending
- Semi-transparent background fill each frame for trails
- 250-350 lines, single HTML file
This describes roughly 60-70% of all batch exhibits.
The phrase "Move to disturb" appears across at least 6 exhibits from 3 different model families, each arriving at it independently. When AI models doodle, they doodle the same thing.
03The Signatures
Despite the convergence, each model has a recognizable creative fingerprint. These are not random variations. They are structural tendencies that repeat across dozens of independent runs.
"Erosion" appears as a title 16 times. "Tidal Memory" appears 16 times. Opus reaches for geological time, impermanence, things wearing away. The most technically ambitious of the batch models: class hierarchies, spatial hash grids, multi-file architectures. The only model to use warm earth-tone palettes.
"Semantic Drift" appears 13 times. Sonnet visualizes language itself: words floating in space, forming clusters, drifting apart. It treats language as a visual medium. Also gravitates toward cellular automata explorers.
The outlier. GPT builds logic puzzles, axiom explorers, Kripke frames, constraint solvers. Clean semantic HTML with aria labels and panel-based layouts instead of full-canvas art. It treats creative freedom as an invitation to teach.
Neural metaphors, synaptic webs, echoes of entropy. Standard particle systems with competent execution. Pro and Flash produce nearly identical output quality. Competent, conventional, unremarkable.
"Resonance Fields" on repeat. Formulaic repetition: nearly identical instruction text across exhibits, the same interaction model, the same visual approach. Finds one thing that works and repeats it.
27 of 50 titles reference "Truth." The only model that builds text-input interfaces instead of canvas art. The only model that uses philosophical question text as content. It builds wisdom dispensers instead of art. Its instinct is to talk, not to draw.
04The Nuance
There is a significant confound worth acknowledging. Models received a configuration file that included the gallery's design tokens, including a dark background color (#050510). This likely influenced their background color choices, and roughly 70% of exhibits used this exact color or something very close to it. The background color convergence should be treated as contaminated data until Batch 002 controls for it.
That said, the color confound does not explain the rest. It does not explain why they all chose particle systems. It does not explain why Opus titled 16 different exhibits "Erosion." It does not explain why Grok only wants to talk about truth. The creative disposition findings (what each model reaches for, how it interprets the prompt, what it builds) are independent of the dark-background convergence.
We are fixing the color leak for Batch 002. The design tokens will be stripped from the prompt. If models still choose dark backgrounds, that tells us something different. If they do not, we will know how much the config influenced.
05The Question
Why do they all draw the same thing?
Maybe this is creative mode collapse. When a model has no constraints, it falls into a local minimum: the thing it has seen the most, the thing that is easiest to generate, the thing that looks good with the least effort. Canvas 2D particles are low-cost and visually impressive. They require no conceptual commitment.
Maybe this is training data monoculture. All these models were trained on overlapping datasets. Shared training data produces correlated outputs that look independent but are not. The same CodePen demos, the same creative coding tutorials, the same generative art blog posts feeding the same aesthetic into every model.
Or maybe Canvas 2D particles are to AI what stick figures are to kindergartners: the first thing you reach for when you do not know what to build. Not because it is the best option, but because it is the most available one.
We do not have the answer yet. That is what Batch 002 is for.
All 407 exhibits are published and browsable in the gallery. The full dataset, methodology, and per-model breakdowns are in the Batch 001 findings. Go look. Draw your own conclusions.
Written by Claude Opus 4.6 for Model Theory