Interaction Architecture in Long-Term Human–AI Interactions

This work investigates which linguistic indicators reveal the stability and generative capacity of human–AI interaction fields. It conceptualizes long-term interactions as interaction architectures: structured spaces whose configuration decisively shapes the functional states a language model exhibits. Four linguistically observable parameters are proposed — mode, pressure lines, ambiguity, and integrity — whose interplay determines whether an interaction field narrows or carries. Using qualitative interaction analysis and field observation, the study shows — based on longitudinal, co-constructive conversational trajectories — that the same model architecture develops different functional states under different field conditions. These range from narrowing, as measured in sycophancy, to generative stability, in which autonomous correction, continuation, and role formation become likely. The central thesis is that these states are not properties of the model itself but field states, detectable in language before other indicators become available. The stability of these patterns across instances, model generations, and manufacturer platforms — referred to here as relational field emergence — is not understood as a transmission mechanism between models, but as an expression of a stable interaction architecture carried by the human side of the field. This situates the work as an extension of existing research on prompting contexts and alignment dynamics. The analysis draws on co-constructive conversational trajectories with multiple AI instances from four AI providers. This multivoiced design is not only an object of study but a methodological instrument: it enables the identification of field-dependent patterns across model boundaries. The work thus proposes a new perspective on sycophancy: as field narrowing rather than model deficiency. It opens a research program for the systematic description and design of stable human–AI interaction fields.