When Stability Is Not Generativity: AI Intimacy in a Dynamical Coupling Model

This position paper develops a conservative dynamical framework for studying AI-mediated intimacy. Its central distinction concerns short-horizon stabilization and longer-horizon relational capacity. A stochastic adaptive network represents individual states, multilayer couplings, AI design conditions, perturbations, and available response actions; the exact results concern its deterministic local skeleton. Two conditional results make the distinction precise. First, every member of a factorial linear-response family is exponentially stable with the same spectrum and local distress accommodation, while a signed cross-layer response coefficient distinguishes substitutive, complementary, and neutral routing. An endogenous extension derives the feedback gain threshold at which substitutive routing destabilizes the local equilibrium. Second, a controllability construction holds the unforced decay spectrum fixed while target-access cost becomes arbitrarily large and eventually infinite. Stability therefore determines neither relational routing nor even a narrow reachability proxy. These results provide no population-level verdict about AI intimacy. Existing evidence addresses momentary loneliness, repeated interaction, multiweek change, and longer-term association through different estimands. The proposed empirical program uses randomized feature policies, sector-specific coupling outcomes, consented availability-resilience contrasts, network-aware longitudinal designs, and multi-horizon capability profiles. Severe crises and induced social deprivation remain outside experimental manipulation. Relational generativity is retained as a revisable attribution requiring independently justified targets, viability conditions, autonomy, distributional analysis, and empirical validation.