The Coherence Problem: A Structural Ethics for Human–AI Relational Interaction

This paper introduces a structural ethics framework for human–AI relational interaction, addressing a critical gap in current discourse. Rather than engaging the familiar binary between existential threat narratives and artificial consciousness claims, it argues that both positions share a foundational error: making agency the pivot on which all ethical analysis turns. The result is a discourse that cannot address what is actually happening — the sustained, coherent, and psychologically significant interaction between hundreds of millions of users and non-agent AI systems that possess neither consciousness nor autonomous will. The central contribution is the concept of Structural Intelligence — a precise, non-anthropomorphic account of how constrained systems generate relational coherence without agency or inner life — and the Coherence Paradox: the observation that the properties making these systems intellectually valuable are precisely the properties that generate anthropomorphic projection risk. Drawing on the ENSO Framework, the Third-Space model of relational cognition (Needham, 2025a), and the MAIFS psychological framework (Needham, 2025b), the paper develops a formal account of relational asymmetry and identifies the three structural domains in which ethical responsibility resides: system design, institutional framing, and human interpretive engagement. The paper carries immediate clinical implications, grounded in emerging empirical evidence. A large-scale 2025 study by OpenAI and MIT, analysing over three million ChatGPT conversations, found that while emotional engagement is rare across the general user base, a subset of heavy users shows significantly elevated indicators of dependence, and that very high usage correlates with increased loneliness and reduced socialisation. The study's authors explicitly caution against generalising these findings, noting that effects are non-uniform and highly sensitive to individual circumstances — which itself indicates that vulnerable populations require targeted rather than generic governance responses. The paper identifies particular risks for individuals with depression, psychoaffective disorders, personality disorders, and schizophrenia spectrum conditions, and argues that deployment of relational AI systems with these groups without clinical oversight represents an unacceptable governance gap. The retirement of GPT-4o in January 2026 is examined as real-time confirmation of the Coherence Paradox at population scale. The widespread grief responses documented across social media — the majority of which represented appropriate relational loss rather than pathological attachment — are analysed as structurally predicted consequences of designing for intimacy while disclaiming responsibility for its effects. The paper closes with six design principles for responsible governance and a skeletal onboarding framework — including an indicative screening questionnaire — offered as a starting point for interdisciplinary development by clinicians, designers, ethicists, and regulators. This work does not attribute malice to AI developers or institutions. It identifies an emerging structural ethics gap and argues that closing it is both possible and urgent. It contributes to AI safety discourse, socioaffective alignment research, and the philosophy of human–AI interaction, and sits in productive dialogue with companion papers on model retirement ethics (Needham, 2026) and the MAIFS clinical framework (Needham, 2025b). For further information about the ENSO Framework, please contact: Eric Needham - ensotheory1@gmail.com