From Agents to Conscious Agents — What Would Count as Evidence for Artificial Consciousness?

Abstract

This working paper develops an evidence framework for distinguishing competent artificial agency from consciousness. It asks two separate questions: when does the evidence justify investigating an artificial system as a consciousness candidate, and what further commitments would be needed to attribute subjective experience to that system? The paper is a philosophical and architectural position paper grounded in a targeted review of philosophy, cognitive science and AI-consciousness research available through 19 September 2026. It distinguishes phenomenal consciousness from access consciousness, agency, self-modelling, metacognition, temporal continuity, individuation, embodiment and valence. Fluent conversation, self-description, memory, personality and humanlike behaviour do not establish experience on their own. The proposed assessment considers six functional dimensions: consequence-sensitive agency; dependence on accessible history; a useful model of the system's own state or limits; functional continuity over time; causal dependence across functions; and monitoring and control of the system's own processing. These dimensions organize inquiry rather than constitute a universal checklist or developmental hierarchy. Their relevance, including any omission, must be justified for the specified theoretical target. A consciousness-candidate designation is a separate, revisable, theory-linked assessment—not a seventh functional capacity or a validated diagnosis. A functional profile does not automatically establish consciousness. Naming a theory is insufficient, and the candidate label itself adds no evidence. The positive case must defend the theory's relevance, its supporting evidence and limitations, and its application to the system. The framework makes these judgments open to scrutiny without providing a validated universal rule for weighting theories or admitting candidates. The evidential discussion addresses convergent behavioural and internal observations, controlled intervention and ablation, strong alternative explanations, recovery, generalization and independent replication. Comparative conclusions remain conditional on the declared alternatives and evaluation conditions. Selective causal effects can support a bounded functional explanation; they do not automatically establish unique realization, universal necessity or subjective experience. Results that support a credible alternative must be allowed to weaken the relevant claim. The theoretical comparison includes functionalism, global workspace and higher-order accounts, recurrent processing, attention schema theory, predictive processing and active inference, embodied and enactive approaches, integrated information theory, and biological naturalism. The problem of other minds and objections concerning absent or inverted qualia remain distinct from engineering performance. Temporal continuity, system individuation, regulatory significance and appropriate epistemic restraint are treated as research questions, without assuming that artificial experience would have to resemble human experience. The contribution is a conceptual and methodological synthesis. No original experiment is reported, no system is awarded consciousness-candidate status, and no positive consciousness or AGI achievement claim is made. The framework and its hypotheses have not been empirically validated; novelty is not established. Reading guide. The English PDF is the authoritative manuscript; the Hungarian PDF is its complete companion translation. Version: v0.2.1. The manuscripts are available under CC BY 4.0; the reused background figure retains its original attribution. Background relation. Material associated with DOI 10.5281/zenodo.22842833 is referenced only as background architecture and the source of a reused figure. This is an independent paper, not a new version of, or appendix to, that material.