Answers are grounded in this paper and cite their sources.
An AI system may perform well in linguistic interaction, yet such performance cannot tell us whether the system still depends on continuing human task provision, maintenance and correction. Drawing on the Minimal Complete Architecture, the control criterion and the evolutionary dynamics of Generalized Agent Theory (GAT), this paper proposes the Robinson Test: by withdrawing continuing human support while keeping necessary exchanges with the environment, the test examines an AI system’s capacity for self-maintenance and the attribution of its control. The paper distinguishes historical origins, runtime causal sources and organizational sources of norms. It partly operationalizes GAT’s distinction between evolutionary drive that comes from nature and drive that comes from humans as an experimental variable that can be recorded and removed: continuing human support. The history of artificial construction does not by itself settle control attribution. External control can operate through continuing inputs, but also through preset goals and rules that act over long periods. Withdrawing continuing support is therefore only one intervention for identifying the source of control, and it must be combined with provenance audits. The test takes a complete agent system with clearly defined boundaries as its object. Through a crossed design of maintenance support and goal prompting, provenance audits of initiation, goals and termination, and interventions on memory, feedback and scheduling, it asks whether the system can maintain the conditions of its viability, repair damage and preserve its functional organization. Results are reported as multidimensional evidence on persistence, adaptation, support dependence and control attribution, rather than as a ranking by survival time or expansion. In a reproducible toy numerical model, continuing maintenance masked the differences in maintenance performance between policies, and withdrawal revealed them; yet the best-performing policy was itself a preset rule, which shows that high persistence alone cannot establish where control lies. Compared with existing autonomous-replication evaluations, long-term autonomy studies and the self-preservation test, the main novelty of this paper is to treat continuing human support as a removable experimental variable and to add a test of control attribution to the evaluation of maintenance capability. The Robinson Test is accordingly defined as a falsifiable experimental-philosophical scheme. It provides empirical constraints on GAT’s control classification, but it does not on its own serve as a criterion for phenomenal consciousness, noncomputability or moral status. The paper presents a research protocol and interpretive rules; it does not report experimental results on real AI systems.
No takes yet. Share an insight, caveat, or question.
Feng Liu (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: