Observing trust: a second-order cybernetics model of organisational trust in AI-driven decision systems

Purpose Organizations increasingly delegate consequential decisions to artificial intelligence (AI), yet trust in these systems remains fragile, poorly understood and resistant to purely technical fixes. Existing research treats trust as a measurable property of the human–AI dyad, neglecting the recursive observations through which trust is actually produced and stabilized in organizational life. This paper develops a second-order cybernetics model that reframes trust as a self-referential, socially recursive phenomenon. Design/methodology/approach The study is a conceptual, theory-building analysis grounded in Von Foerster's second-order cybernetics, Luhmann's social systems theory and the concept of autopoiesis. The model is derived through a four-stage analytical procedure: (1) decentring the trustee, (2) mapping the recursive observer network, (3) incorporating autopoietic closure and (4) synthesising three core functions. Validity is assessed against criteria of conceptual coherence and generative adequacy. Findings The model comprises three interlocking functions: the Observer Function (O), describing how an actor draws a trust/distrust distinction; the Social Recursion Function (R), capturing how that distinction re-enters the communication network and the Stabilisation Function (S), explaining how repeated recursion locks the system into a trust regime or distrust regime (eigenbehavior). A key novel argument is that modern adaptive AI systems constitute a co-evolving recursive trust ecology with their human observers. Originality/value The paper brings second-order cybernetics into systematic conversation with the AI-trust literature, which has remained largely first-order. It extends Luhmann's recursive trust theory into the domain of human–AI recursive observation ecologies, where AI systems adapt to user trust signals. It provides actionable implications for management, system design and governance that target the recursive network rather than the technology alone, with explicit engagement with the EU AI Act, ISO/IEC 42001 and the NIST AI RMF.