One Interaction, Three Fields: The Disciplinary Phase Transition Driven by AI Foundation Models

(Part 2 of a 3-part series.) (Revised version (March 2026): expanded experiments including ProtBERT validation, GPT-2 negative control, and stratified long-range analysis) When any field adopts AI methodology for data-driven analysis, it passes through the same computational pipeline. This paper argues that this passage makes methodological convergence across disciplines effectively inevitable—not as a logical necessity, but because the pipeline consistently matches or exceeds domain-specific methods, leaving fields no competitive reason to resist adoption. It demonstrates this inevitability in its strongest form: across biology, chemistry, and physics—the fields with the most entrenched domain-specific methods. We introduce the Learned Interaction Pipeline—a four-stage computational structure (tokenize, interact, aggregate, decode)—and show that it is not specific to any domain but inherent to all AI-based data analysis. Four independently developed models—AlphaFold (biology), MACE-MP-0 (chemistry), the General Physics Transformer and Fourier Neural Operator (physics)—each instantiate this pipeline, differing only in domain-specific implementation choices. That these fields adopted the same computational structure is not surprising; it is inevitable. The pipeline's generality is not an objection but the mechanism: any field that adopts AI-based analysis must pass through this structure, and passage through a shared structure produces methodological convergence. This paper demonstrates this inevitability in three ways. First, structurally: the four models map onto the same pipeline across tokenization, interaction, aggregation, and decoding stages. Second, empirically: when a chemistry model (MACE-MP-0) and a biology model (ESM-2) process the same molecules, their internal representations converge at interaction layers and diverge at output layers—a pattern reproduced across 27 proteins spanning 10 to 164 residues (p = 0.0001, permutation test; p = 2.3 × 10⁻¹², paired t-test). The convergence depends on physical content: randomizing atomic coordinates (p = 1.0 × 10⁻¹³) or shuffling amino acid sequences (p = 6.6 × 10⁻¹¹) each significantly reduces cross-model similarity. Third, functionally: the chemistry model's representations predict protein residue contacts above chance (mean AUC ≈ 0.58) despite having never been trained on biological data, confirming that the shared pipeline carries cross-disciplinary information when trained on physical data. A critical control establishes the boundary of this effect: a general-purpose language model (GPT-2) shows no meaningful representational convergence with either model (mean CKA 0.07 vs. 0.27 for MACE–ESM-2), and its apparent contact prediction is an artifact of positional encoding rather than sequence content (random letter strings perform equally, p = 0.22). Pipeline sharing alone is insufficient; convergence requires that the pipeline be trained on data from the natural world. Domain-specific training does not erase this convergence but modulates it: training on chemistry data reduces cross-model similarity in 15 of 20 proteins tested, with the effect absent below 23 residues and consistently present above 28—the scale at which emergent structural properties begin to dominate. This is Anderson's "more is different" operating in representation space: the shared pipeline dissolves the methodological boundary, while domain-specific learning rebuilds the intellectual distinctness of each field at the scale where emergence begins. The methodology has melted; what remains between disciplines is not a difference of tools but a difference of questions. Because AI methodology drove this dissolution, the pace of progress across these fields is now structurally coupled to the pace of AI advancement.