From Incentive Compatibility to Justified Symbiosis: The Normative Foundations of Symbiotic Intelligence

Abstract Long-term interest compatibility between artificial intelligence and humans is an important condition for human–AI symbiosis, but it is not a sufficient condition for justification. An AI system may form a stable, mutually beneficial relationship with its direct users while weakening their independent decision-making capacity, externalizing costs to third parties, creating barriers to exit, or allowing responsibility to disappear across a complex deployment chain. Incentive compatibility characterizes "which behaviors are sustainable under a given payoff structure"; a normative theory must answer "which sustainable relationships are worth allowing, how conflicts should be adjudicated, and how baselines should enter system design." This paper proposes the framework of **Symbiotic Normativity**. Its core contribution is not another list of AI ethics principles, but the transformation of five principles into normative predicates that are decidable at the level of actions, institutions, and history: human agency, genuine reciprocity, multi-party fairness, right to exit, and accountability. Normatively acceptable actions are the joint feasible set of institutionally feasible actions and the five predicates, rather than a conversion of norms into tradable weights within a utility function. The main line of justification for the five principles is that they are normative mirrors of five failure modes of incentive-compatible relationships: dependency manufacture corresponds to Agency, unilateral extraction to Reciprocity, third-party exclusion to Fairness, lock-in to Exit, and responsibility dilution to Accountability. This enumeration is not an axiomatic derivation but a refutable, reflective-equilibrium justification. At the mechanism level, the paper provides a five-step layered adjudication procedure with actors, powers, deadlines, budgets, and consequences, and it explicitly accepts the trade-off that lexicographic baseline filtering entails for social-choice axioms. The paper further offers a **design proposition** for the norm–incentive interface: in a self-contained minimal game, if manipulative actions are institutionally excluded and, after exclusion, the n Bilingual edition: English (en-v1.0) and Chinese (zh-v1.0) PDFs are included in this record.