Do Different AI Models Agree on What Symbols Mean, and Why? A Proposed Study Design on Cross-Model Semantic Convergence for Visual Symbols

This research proposal asks whether vision-language models from different developers converge on the meanings of visual symbols, and what such convergence could establish. The design separates visual recognition, conventional association, and abstract interpretation. Rather than treating agreement on familiar symbols as sufficient evidence, it focuses on shared incorrect answers to obscure signs and semantic convergence on privately generated, never-published glyphs. Proposed methods include difficulty-graded stimuli, controlled visual manipulations with matched sham controls, human familiarity and identification baselines, model-provenance records, and convergence measures calibrated against shuffled baselines and within-model repeatability. Separate exposure and intervention arms examine documented training corpora, convention acquisition in synthetic micro-worlds, and the effort required to overwrite existing associations. The proposal distinguishes behavioural agreement from causal evidence: convergence alone cannot establish that models share training data or that a meaning is inherent in an image. Status: proposed study only. Nothing has been run or piloted; no data or empirical results exist, and the proposal has not been peer reviewed. Outstanding methodological decisions and limitations are explicitly documented. The design is offered as a starting protocol for researchers or groups equipped to implement, refine, or adapt it. Concept: Kon Lionis. Drafted with Claude (Anthropic), October 2026.