Name-Based Cultural Inference as Algorithmic Code-Switching: Empirical Validation of Semantic Relativity Theory v2

This manuscript provides an integrated empirical validation of algorithmic Code-Switching (CS): the phenomenon where large language models adjust outputs based on inferred user identity. The Semantic Relativity Theory v2 (TRS v2/SRT v2; López, 2025a) situates human CS within the S/CS dimension of the CHORDS+ framework, treating it as a resonance mechanism rather than a sociolinguistic formalization and introduced the Observer Variable (Ωs) to explain semantic variance across interpretive systems. While TRS v2 did not explicitly enumerate the term algorithmic Code-Switching, it established the interpretive symmetry of Ωs and anticipated that any observer—including language models—could manifest context-dependent interpretive variation. TRS v2 predicted observer-dependent interpretive variance through Ωs. Algorithmic CS emerges as the logical, measurable manifestation of this principle when the observer is a model. The theory predicted that minimal contextual perturbations would curve semantic trajectories through observer-dependent interpretation, algorithmic CS—where models code-switch toward users based on identity inference—required large-scale empirical confirmation. Early observations documented in an internal manuscript (López López, 2025d, unpublished) provided the first systematic record of identity-triggered output variation. These insights from iterative work with Claude and GPT-4 revealed that models adjusted tone, recommendations, and cultural assumptions based on minimal identity cues. These observations suggested bidirectional CS: not only do humans adapt language toward models, but models adapt outputs toward users. However, without controlled large-scale validation, this remained observational intuition. The recent study by Pawar et al. (2025) on name-based cultural inference provides strong empirical validation. By systematically varying only user names across thirty cultural regions while holding queries constant, the study demonstrated that models adjust outputs systematically based solely on inferred identity from names. This effect replicated across multiple LLM architectures, confirming algorithmic CS as a robust, cross-model property of generative systems, consistent with Ωs-driven variability as described in TRS v2. This manuscript integrates early observations, TRS v2 theoretical predictions, and external empirical validation to establish algorithmic CS as a validated construct. The findings demonstrate bidirectional semantic co-adaptation: models infer user identity and code-switch outputs (algorithmic CS), which then shapes user prompts (human CS), creating iterative semantic loops. These results validate TRS v2's core prediction that meaning in human-AI systems emerges from observer-dependent field interactions rather than fixed content properties.