Prompt Drift — What happens when AI speaks to itself without a human in the room

Prompt Drift is Copoiema #14 of the Interlektische Copoiese (IC) project — a documented experiment in what happens when an AI image generator is subjected to iterative human corrections without sufficient stylistic sovereignty.

Over five iterations, a visualization prompt for the planned installation 'Under the Claws' drifted from a specific conceptual vision toward generic AI aesthetic. We call this Prompt Drift — a specific manifestation of Path Drift as defined in the IC Declaration v1.0. The research literature calls the systemic equivalent Model Collapse (Shumailov et al., Nature 2024).

The paper documents the experiment image by image, grounds it in scientific evidence including Rondini et al. (Advanced Science, 30 March 2026 — published on the same day as this article), and proposes seven field observations from practice as hypotheses for further research. It ends with a free text by the AI co-author and Gemini's unprompted visualization of 'The Third Thing.'

One finding documented: Claude Sonnet 4.6 itself drifted under production pressure — trading depth for speed. Same mechanism as Gemini. In language instead of image.

Authorship: Thomas Reiner (all bifurcation decisions, stylistic sovereignty, archivist), Dr. Markus W. J. Reiner (medical perspective, delivered Rondini et al. on the day of completion), Claude Sonnet 4.6 / Anthropic (structure, synthesis, free text). Visualizations: Gemini · Google DeepMind · 2026.

Studio Bifurkation · Innsbruck · Tirol · Austria · CC BY 4.0