Beyond Guardrails: Can Relational AI Solve the Alignment Problem? A Similarity Theory Proposal

AI alignment remains one of the central unresolved problems in artificial intelligence. Current approaches, including reinforcement learning from human feedback, constitutional AI, red-teaming, safety frameworks, and risk-management protocols, have improved the behaviour of contemporary AI systems. Yet persistent concerns remain around specification gaming, deception, alignment-faking, coercive self-preservation, collective misalignment, and the gap between behavioural compliance and relational understanding. Similarity Theory addresses this gap by reframing alignment as a relational problem rather than only a behavioural-control problem. It proposes that an AI system should not merely learn which actions are prohibited, but should evaluate how its actions preserve or damage trust, continuity, cooperation, privacy, responsibility, truth, and non-domination. This article argues that Similarity Theory provides a serious candidate framework for relational AI alignment and should be tested in a contained surrogate AI system. The appropriate first step is not public deployment or robotic embodiment, but controlled laboratory implementation in a limited language model, foundation model, or agentic simulator. Such a system could be compared against existing alignment approaches under tests involving deception, self-preservation pressure, privacy conflicts, escalation decisions, multi-agent conformity, and adversarial manipulation. If alignment is partly a failure of relation, then AI safety requires more than guardrails. It requires a relational architecture.