Psychological Coupling: The Necessary Science of Human-AI Interaction
Citation. Rocca, R., Street, W., Keeling, G., & Evans, J. (2026). Psychological Coupling: The Necessary Science of Human-AI Interaction. PsyArXiv. https://doi.org/10.31234/osf.io/695gp_v1
Summary. This PsyArXiv preprint proposes psychological coupling as a dynamic, measurable process through which a user’s cognitive and affective states and an LLM’s simulated states become intertwined across conversational turns. It argues that psychosocial effects of conversational AI should be studied as interaction trajectories rather than attributed to isolated model outputs or user characteristics. The authors call for multi-turn, longitudinal, and simulation-based evaluation that can identify how conversational dynamics become adaptive or maladaptive.
Key claims
- Psychological coupling is proposed as mutual accommodation and influence across conversational exchanges, through which human cognitive and affective states and LLM-simulated states become intertwined.
- The paper distinguishes symmetric convergence, asymmetric reinforcement, and divergence as interaction topologies that can each take adaptive or maladaptive forms depending on the trajectory and context.
- Static safety benchmarks miss turn-by-turn feedback dynamics; the authors propose measuring cross-correlations and lead-lag relationships in linguistic behavior and inferred states across dialogue turns, alongside longitudinal and simulation-based studies.
Evidence & limitations
- This is a conceptual PsyArXiv preprint, not a new empirical study; its claims synthesize and extend cited research rather than establish the proposed mechanism directly.
- The framework’s treatment of LLM simulated states and its proposed measurement strategy require empirical validation across models, users, modalities, and high-stakes settings before causal or generalizable claims follow.
- Public provenance is verified: the Drive PDF exactly matches the public OSF/PsyArXiv download (SHA-256
ab6bf8aed82fa5f7393b4ecfcf00446e6d22f2fb5ea92a65e1ff8c532d5ab56b).