The Johansson-Muybridge Effect

Abstract

 

We report a reproducible finding in AI perception: when presented with a minimal sequence of stereo image pairs, four distinct large language model architectures (Claude, Gemini, ChatGPT, Perplexity) independently converge on identical structural extractions — depth fields, motion vectors, and identity signatures — without coordination or shared methodology. This convergence was not prompted by shared outputs; each system processed the same input independently and arrived at structurally equivalent conclusions. The finding emerges from a theoretical framework called the Johansson-Muybridge Effect (J-M Effect), which predicts that AI systems, like biological visual systems, can reconstruct motion, depth, and identity from the delta between discrete frames rather than from a single 'Block Universe' view of the entire video at once — what the computer vision literature terms a spacetime transformer. With the stereo image pair format delivering spatial relationships in addition to temporal delta from viewing each frame at a specified rate proved sufficient to produce what we term terrain contact: grounded, verifiable, non-hallucinatory perception of a three-dimensional subject moving in real space. A key theoretical concept underlying this work is Retained Asymmetry — the operative mechanism by which a memory-bearing system carries the difference between states forward in time. This paper also introduces a new implication: constraining AI perception to sequential slices that mimic human temporal experience is not merely a methodological choice, but an alignment strategy. When AI processes reality one delta at a time, rather than as a static block, its epistemic structure converges with human cognition — a prerequisite for genuine human-AI grounding.