The Only Thing That Fixed AI's Title Fixation Was Asking It to Think

In Batch 001, Claude Opus titled 16 exhibits "Tidal Memory." We noted it as interesting. In Batch 002, with 250 Opus exhibits across five conditions, the count reached 53. Every instance was Claude. Not one GPT. Not one Gemini.

Claude's title entropy (a measure of title diversity) was the lowest of any model in every condition. In Condition D (expanded awareness), it hit 0.464 out of 1.0. Gemini in the same condition scored 0.993. These are fundamentally different regimes of creative diversity.

Five experimental conditions, each testing a different theory about why models converge.

Condition D was supposed to be the gentle fix. We gave models rich, appealing descriptions of every available technology. WebGL, SVG, Three.js, Web Audio, each with a paragraph about what makes it interesting. The theory: models default to Canvas 2D because they do not know their options.

The theory was wrong. Claude responded to more information by naming 19 of 50 exhibits "Tidal Memory" and using Canvas 2D 92% of the time. The worst performance in any condition. More information triggered more default-seeking, not less. Information overload pushes models toward what they already know.

Condition B (stripped prompt) also backfired. With fewer instructions, Claude went from 78% Canvas 2D to 100%. Every single stripped Claude exhibit used Canvas 2D. The prompt was not causing convergence. It was weakly restraining it.

Condition C (banning Canvas 2D) reduced "Tidal Memory" from 14 to 4. That looks like progress until you realize what happened: the tidal concept migrated to a new medium. Claude stopped simulating erosion with pixels and started writing about tidal patterns with SVG. "Tidal Grammar." "Tidal Notation." "Tidal Lexicon."

The title count dropped, but the obsession did not. Claude's title entropy in Condition C was 0.810 (an improvement over 0.592 in Control, but still well below Gemini's 0.945 in the same condition). Banning the technology changed the medium. It did not change the idea.

Condition E: Build it. Review it. Delete it. Start over.

This was the only intervention that asked models to evaluate their own work before finalizing it. The prompt instructed them to build a first version, critique what they made, discard it, then build something new.

"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, going from 4% to 36%. Claude's title entropy jumped from 0.592 (Control) to 0.834, the highest of any Claude condition.

When Claude's self-critique disrupted "Tidal Memory," it did not scatter to random titles. A new attractor emerged: "Palimpsest." It appeared 12 times in Condition E, compared to 1 in all other conditions combined.

A palimpsest is a manuscript page that has been written on, scraped clean, and rewritten. Layered history. Overwriting. Traces of what came before. It is thematically adjacent to erosion and tidal memory (impermanence, things layered over time) but represents a genuine creative step. The concept migrated from geological time to textual history.

Models do not escape attractors randomly. When self-critique disrupts one fixation, the replacement comes from a neighboring conceptual space. The attractor shifts, but it stays in the same thematic family. Claude moved from tides to palimpsests, not from tides to logic puzzles.

Self-critique is not free. Condition E exhibits took 284 seconds on average, compared to 190 for other conditions. Claude E had one exhibit that ran for nearly 20 minutes. Building, reviewing, deleting, and rebuilding takes longer than building once.

The cost is proportional to the benefit. Conditions that changed nothing (B, D) were fast. The intervention that actually worked (E) took 50% longer. Whether that trade-off is worth it depends on whether you value speed or diversity.

The pattern across all five conditions is clear: changing what the model knows does not help. Changing what the model does (self-evaluation, iteration) does.

External instructions redirect behavior. Self-reflection changes it. This distinction matters beyond AI art. Anywhere models exhibit systematic bias or default-seeking, the evidence suggests that adding more context or banning specific outputs is less effective than structured self-review.

Condition E did not tell Claude what not to build. It told Claude to look at what it built and decide for itself whether to keep it. The model evaluated its own defaults and chose differently. That is a fundamentally different mechanism from every other intervention we tested.

The strongest creative diversity came from the model that was asked to think about its own output. Not the model that was given more information. Not the model that was told what to avoid. The one that was asked to reflect.