Telling AI Not to Draw Circles Made It Draw Something Else

Condition C was the blunt-force intervention. The prompt included an explicit instruction: do not use Canvas 2D for rendering. Do not use dark backgrounds. Everything else remained the same. Same creative freedom. Same sandbox. Same isolation protocol.

150 exhibits ran under this constraint, 50 per model (Claude Opus 4.6, GPT 5.2, Gemini 3 Pro). The question was simple: if you remove the default, what fills the gap?

The technology shift was total.

Canvas 2D dropped from 50.7% to 1.3%. The two that remained were Claude exhibits that ignored the instruction. All 50 GPT and 48 of 50 Gemini exhibits complied perfectly.

SVG became the dominant medium, jumping from nonexistent to 67.3%. Models that had never touched SVG in 407 previous exhibits suddenly produced hundreds of lines of path data, viewBox calculations, and SMIL animations. Three.js went from 1% to 12%. CSS animations went from a rounding error to more than a quarter of all output.

The visual character of Condition C is immediately recognizable when browsing the gallery. Light backgrounds. Clean lines. Geometric shapes instead of particle clouds. The warm, dusty palettes that replaced the universal dark monochromes are striking.

Claude's SVG output leans toward organic curves: wave forms, flowing paths, shapes that recall tidal motion (more on that in a moment). GPT built structured layouts with labeled components, turning SVG into diagrammatic tools. Gemini produced the most visually diverse set: crystalline geometries, recursive patterns, color gradients.

The average line count dropped from 572 (Control) to 507 (Anti-Default). SVG is more concise than Canvas 2D. A single SVG path element replaces dozens of lines of requestAnimationFrame particle logic.

This is the finding that surprised us. Banning a technology changed what models wanted to say, not just how they said it.

Claude stopped building erosion simulations. Erosion requires pixel-level Canvas control: drawing individual grains, simulating water flow across a surface, alpha-blending sediment layers. Without Canvas, the concept does not work. "Erosion" dropped from 9 appearances in Control to just 1 in Anti-Default.

GPT showed the same pattern. In Control, 48% of GPT exhibits were model-theory themed (axiom explorers, back-and-forth games, structure finders). In Anti-Default, that dropped to 12%. GPT's teaching tools depend on panel layouts with interactive controls. Without its preferred medium, GPT pivoted to generative SVG art entirely.

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.

Claude's adaptation to the Canvas ban is the clearest demonstration of concept-medium coupling. "Tidal Memory" appeared 4 times in Condition C (down from 14 in Control). But a new pattern emerged: tidal variants.

This is not a random title change. "Tidal Grammar" and "Tidal Notation" are linguistic concepts applied to the tidal framework. Claude lost access to pixel-level water simulation, so the tidal concept migrated from visual simulation to linguistic abstraction. The obsession adapted to the constraint.

Condition C tested whether AI creative convergence is a technology problem or a conceptual one. The answer is both, but they are not independent.

Banning Canvas 2D eliminated the visual archetype (dark particle systems). It forced genuine technology diversification. It produced lighter, more varied visual output. In those ways, it worked.

But it did not eliminate thematic fixation. Claude's tidal obsession migrated to SVG. GPT's medium shift was so complete that it lost its creative identity entirely. The intervention changed the surface but not the structure.

Compare this to Condition E (forced self-critique), which actually disrupted thematic attractors. Banning defaults redirects behavior. Self-reflection changes it. These are different mechanisms, and both matter. But if you want a model to genuinely explore new creative territory, telling it what not to do is less effective than making it evaluate what it already did.