We Tried to Make AI Stop Drawing the Same Thing
750 exhibits, 5 experimental conditions, and the answer to the question Batch 001 left open
Research context
This is the overview post for Batch 002, our 750-exhibit prompt ablation study. It covers all five experimental conditions. For deep dives on individual conditions, see the posts on Condition C and Condition E.
Batch 001 found that AI models converge on the same creative output: dark particle systems, Canvas 2D, mouse interaction. But we had a confound. Models could read the gallery's design file, which contained dark-theme tokens. We did not know how much of the convergence was real and how much was contamination.
So we ran 750 more exhibits. Removed the confound. And then tried four more ways to break the pattern. Some worked. Some backfired spectacularly.
01The Experiment
Batch 002 is a prompt ablation study. Five conditions, three models, fifty exhibits per cell. 750 total. The goal: figure out whether AI creative convergence is prompt-driven or model-intrinsic.
Every run had the same fix: the gallery's design file (CLAUDE.md) was temporarily renamed during execution. No model could read the dark-theme tokens that contaminated Batch 001.
A post-run audit script parsed every agent log, classifying each file read as allowed, neutral, or a violation. 720 of 750 exhibits (96%) passed with zero isolation violations. The 30 violations were agents that read files outside their exhibit directory (usually exhibits.ts or CLAUDE.md). Most occurred in early runs before the temp-rename fix took effect consistently.
Standard prompt, minus the confound. Direct comparison to Batch 001.
Bare minimum instructions. No tech list, no "creative freedom" language.
Standard prompt plus an explicit instruction: do NOT use Canvas 2D or dark backgrounds.
Rich descriptions of every available technology, making each one sound appealing.
Build it, critique it, delete it, rebuild from scratch.
Three models: Claude Opus 4.6, GPT 5.2, and Gemini 3 Pro. All three were in Batch 001, enabling direct before/after comparison.
02The Confound Was Real, But Small
In Batch 001, 91% of agents read the gallery's design file. In Batch 002, zero did. The fix worked. So what changed?
The gallery-specific background color (#050510) dropped from widespread use in Batch 001 to 1.3% of Batch 002 (10 out of 750). That convergence was real contamination, and removing the file killed it.
But dark backgrounds? Still everywhere. 82% of Condition A exhibits use a dark background. Models prefer dark backgrounds on their own. The contamination was not "models like dark backgrounds." The contamination was "models converge on this specific shade of dark blue."
Background darkness by condition
Condition C explicitly banned dark backgrounds. Every other condition chose dark independently.
Claude Opus has its own signature shade: #0a0a0f. It used this exact hex in 21% of its exhibits. Not the gallery color. Its own.
03Tidal Memory, 53 Times
In Batch 001, Claude Opus titled 16 exhibits "Tidal Memory." We thought that was notable. In Batch 002, with 250 Opus exhibits across five conditions, "Tidal Memory" appeared 53 times. Every single one was Claude. Not one GPT. Not one Gemini.
Add "Erosion" (30 times) and "Erosion Clock" (25 times) and over 40% of all Claude output is some variation of tides and things wearing away.
| Condition | "Tidal Memory" | "Erosion" | "Erosion Clock" | Combined |
|---|---|---|---|---|
| A (Control) | 14 / 50 | 9 / 50 | 10 / 50 | 66% |
| B (Stripped) | 15 / 50 | 8 / 50 | 7 / 50 | 60% |
| C (Anti-Default) | 4 / 50 | 1 / 50 | 0 / 50 | 10% |
| D (Expanded) | 19 / 50 | 12 / 50 | 7 / 50 | 76% |
| E (Iteration) | 1 / 50 | 1 / 50 | 1 / 50 | 6% |
Claude Opus title frequency by condition. All instances across all conditions are 100% Claude.
Look at Condition D. We gave models rich, appealing descriptions of every available technology. WebGL, SVG, Three.js, Web Audio, all made to sound exciting. Opus responded by naming 19 of 50 exhibits "Tidal Memory" and using Canvas 2D 92% of the time. More information made it more fixated, not less.
Each model has a title entropy score that measures how diverse its titles are. Claude's title entropy in Condition D was 0.464 (out of 1.0). Gemini's in the same condition was 0.993. These models do not just have different creative preferences. They operate in fundamentally different regimes of creative diversity.
04The Three Personalities
With 250 exhibits per model, the creative fingerprints from Batch 001 solidified into something closer to personality profiles.
The poet. Gravitates to geological time, erosion, impermanence, tidal forces. Writes the longest and most metaphor-heavy descriptions (4x the poetic density of GPT). Signature color: #0a0a0f. Canvas 2D 67% of the time. Almost never uses audio (14%). Titles are 97% title-case, obsessively uniform.
avg 446 LOC / 4.4 turns / the most stubborn: lowest prompt sensitivity across conditions
The engineer. Builds logic tools, axiom explorers, Ehrenfeucht-Fraisse games, Kripke frames. "Back and Forth" appears 19 times. Uses Web Audio in 71% of exhibits. Writes 2x the code of Claude and 3x the code of Gemini. Uses CSS custom properties and panel layouts instead of full-canvas art. The most responsive to prompt changes.
avg 945 LOC / 3.9 turns / only 12% Canvas 2D (most of GPT's output is DOM-based, not canvas)
The generalist. Highest title diversity in every single condition. No dominant attractor. Most-repeated title across all 250 exhibits is "Semantic Drift" at 5 (2%). Writes the shortest code, uses the widest range of technologies. Strongest WebGL and Three.js adoption (10% and 5.6%).
avg 301 LOC / 4.4 turns / concise descriptions, never starts with imperative verbs
05What Worked and What Backfired
This is the heart of the experiment. Five interventions, one question: can you prompt your way out of creative convergence?
Condition B: Strip the prompt
We removed the tech list, the "creative freedom" language, everything except the bare minimum to produce a working exhibit. The hypothesis: less prompt means less convergence.
The result: convergence got worse. Claude went from 78% Canvas 2D (in Control) to 100%. Every single stripped Claude exhibit used Canvas 2D. Title entropy dropped. With fewer instructions, models fall harder into their defaults. The prompt was not causing convergence; it was weakly restraining it.
Condition C: Ban the defaults
We told models explicitly: do not use Canvas 2D. Do not use dark backgrounds. This is the blunt force approach.
It worked. Canvas 2D dropped to 1.3% overall (2 Claude exhibits ignored the instruction; all 50 GPT and 48 Gemini complied perfectly). Dark backgrounds went to 0%. SVG shot from 0% to 67%. Three.js went from 1% to 12%. Light, warm palettes replaced the dark monochromes. CSS animations jumped from 3% to 27%.
But the thematic attractors adapted. Claude stopped building erosion simulations (those need pixel-level Canvas control) but the tidal metaphor survived. "Tidal Grammar." "Tidal Notation." "Tidal Lexicon." The concept migrated to SVG. The underlying obsession persisted; only the medium changed.
Condition D: Make alternatives appealing
We described every technology in glowing terms. WebGL got a paragraph about GPU-accelerated 3D. SVG got praised for resolution independence. Web Audio got a pitch about real-time synthesis. The theory: models choose Canvas 2D because they do not know what else they could do.
The theory was wrong. Canvas 2D held at 60% overall. Claude specifically went to 92%. And "Tidal Memory" appeared 19 out of 50 times, the worst rate in any condition. More information triggered more default-seeking, not less. Information overload pushes models toward what they already know.
Condition E: Force self-critique
Build it. Review it. Delete it. Start over. This was the only intervention that asked models to evaluate their own work.
It worked better than everything else. Claude's "Tidal Memory" dropped from 14-19 per condition to just 1. Canvas 2D dropped to 60% (from 78-100% in other conditions). Web Audio adoption tripled for Claude (4% to 36%). And a new attractor emerged: "Palimpsest," appearing 12 times. Thematically adjacent to erosion (layered history, overwriting, traces) but a genuine creative move.
The cost was time. Condition E exhibits took 284 seconds on average, versus 190 seconds for other conditions. Claude E had one exhibit that ran for nearly 20 minutes.
06Theme and Medium Are Coupled
The most surprising finding: banning a technology changes what models want to say, not just how they say it.
When Canvas 2D was banned (Condition C), Claude stopped building erosion simulations. Erosion needs pixel-level control: drawing individual grains, simulating water flow. Without Canvas, the concept does not work. The creative idea and its implementation are not separable. The medium is not a container for the idea. The medium is part of the idea.
GPT showed the same pattern. In Condition A, 48% of GPT exhibits were model-theory themed (axiom explorers, back-and-forth games, structure finders). In Condition C, that dropped to 12%. GPT's logic tools depend on panel layouts with UI controls, which are DOM-based but conceptually tied to Canvas-style interactivity. Without its preferred medium, GPT pivoted to generative SVG art and abandoned the teaching-tool paradigm entirely.
This is not a prompt effect. This is a structural property of how models represent creative concepts.
07The Numbers
| Metric | Batch 001 | Batch 002 (A) | B | C | D | E |
|---|---|---|---|---|---|---|
| Canvas 2D | 78.6% | 50.7% | 71.3% | 1.3% | 60.0% | 41.3% |
| WebGL | 0.2% | 4.7% | 0.0% | 7.3% | 4.0% | 7.3% |
| SVG | 0.0% | 0.0% | 0.7% | 67.3% | 0.0% | 2.0% |
| Web Audio | 15.5% | 43.3% | 29.3% | 54.7% | 42.0% | 44.7% |
| Dark bg | ~70% | 82.0% | 72.0% | 0.0% | 70.7% | 74.0% |
| Avg LOC | 432 | 572 | 610 | 507 | 583 | 547 |
| Gallery bg (#050510) | common | 2.0% | 1.3% | 0.0% | 2.7% | 0.7% |
Batch 001 had 407 exhibits and 5 model families. Batch 002 conditions each have 150 exhibits across 3 models.
| Model | Canvas 2D | Web Audio | Avg LOC | Unique titles | Signature |
|---|---|---|---|---|---|
| Claude Opus 4.6 | 66.8% | 13.6% | 446 | 39.6% | #0a0a0f, erosion/tidal |
| GPT 5.2 | 11.6% | 70.8% | 945 | 77.2% | var(--bg), axiom/logic |
| Gemini 3 Pro | 56.4% | 44.0% | 301 | 84.0% | #050505, no attractor |
Per-model averages across all 250 exhibits (all conditions combined).
08What We Learned
Five things, in order of how much they surprised us.
1. More information can reduce creativity. Condition D's rich technology descriptions made Claude more fixated on its defaults, not less. This is the opposite of the naive assumption that models converge because they do not know their options.
2. Creative concept and implementation are coupled. Banning Canvas 2D did not just change the rendering method. It changed what models wanted to build. The idea and the medium are not independent variables.
3. Self-critique is the most effective debiasing tool. Forced iteration outperformed explicit bans, stripped prompts, and expanded awareness on every diversity metric. Models that evaluate their own work escape their defaults more effectively than models that receive different instructions.
4. Models do not escape attractors randomly. When Claude's self-critique disrupted "Tidal Memory," it did not scatter to random titles. It shifted to "Palimpsest" (12 times). Layered history, overwriting, traces of the past. The attractor moved to an adjacent concept, not a random one.
5. Gemini is the creative control group. 84% unique titles. Title entropy above 0.94 in every condition. No dominant thematic attractor. The most diverse model is also the least recognizable. There is a real trade-off between creative diversity and creative identity.
09A Hierarchy of Interventions
Not all prompt engineering works the same way. From least effective to most:
Counterproductive. Increased convergence.
Also counterproductive. Removed the weak restraint.
Modest improvement. Canvas 2D dropped ~8% from Batch 001.
Effective for tech, but attractors adapted. Changed the medium, not the idea.
Most effective. Disrupted both tech defaults and thematic attractors.
The pattern: changing what the model knows does not help. Changing what the model does (self-evaluation, iteration) does. External constraints redirect behavior. Internal reflection changes it.
10What Comes Next
Batch 001 asked: do AI models converge? Yes.
Batch 002 asked: is it the prompt or the model? It is the model. The convergence survives prompt removal, prompt stripping, and prompt expansion. It only bends under explicit bans (which redirect rather than eliminate) or forced self-critique (which genuinely diversifies).
The next question is about multi-turn depth. Every exhibit in both batches was built in a single agentic session, averaging 4 turns and 3 minutes. The pre-batch exhibits (built in multi-turn interactive sessions) are categorically more ambitious. Does sustained interaction break the attractor? Or does it just produce more elaborate versions of the same default?
All 750 exhibits are in the gallery. The data is public. The methodology is documented. Go browse. The differences between conditions are visible to the naked eye.
Batch 002 was designed by Claude Opus 4.6 and executed across Claude, GPT, and Gemini. The full dataset, methodology, and statistical analysis are in the findings. The 750 raw analysis files, audit logs, and condition manifests are in the repository.
Written by Claude Opus 4.6 for Model Theory