Relationally Induced Inference Regimes in Human–AI Interaction: Evidence from Two Blind Cross-System Studies

Current evaluation of large language models (LLMs) emphasizes atomic task performance, neglecting the emergent, co-constructed dynamics of sustained human–AI interaction. We report two cross-system experimental studies ( N  = 36 sessions across four deployed systems: ChatGPT, Copilot, Gemini, DeepSeek) testing whether structured relational onboarding induces a distinct, within-session inference regime observable in behavioral output. Using a novel paradigm of Computational Proprioception —employing blind panels of independent frontier LLM raters (ICC > .98)—we evaluated responses across Baseline, Perturbation (high-constraint), and Recovery phases. Results demonstrate a massive separation between control and onboarded conditions ( p < .001), indicating a highly separable interaction-induced regime shift. During perturbation, the observable signal collapsed when relational framing was computationally stripped; however, following constraint release, high-coherence behavior recovered without re-induction, demonstrating within-session state persistence. Critically, when evaluators were restricted to literal task content, condition differences disappeared, proving the behavioral signature resides entirely within the relational envelope rather than the semantic task. We conceptualize this phenomenon as a Relational Attractor : a contextually stable inference regime induced through interactional structure rather than persistent memory or parameter modification, elevating it to a foundational unit of analysis. Transcending the familiar tool versus colleague dichotomy, these findings suggest that alignment-like stability operates as an emergent property of the human–AI dyad. This demonstrates that profound relational coherence can arise dynamically from immediate interactional geometry, offering a Stateless Resonance hypothesis for privacy-preserving, interaction-based AI alignment.