<p><?tk 2?>Since Turing’s imitation game, artificial intelligence has been evaluated primarily through intelligence—whether a machine can think, understand, or reproduce human cognition. Yet intelligence is conceptually unstable even in human cases. This paper argues that, while intelligence remains valuable for technical progress, it is too narrow as a societal benchmark because it fixes attention on component-level performance while the central moral stakes arise at the level of sociotechnical deployment: who remains answerable, how accountability is allocated, and whether redress is structurally available. Methodologically, the paper separates two analytic axes commonly collapsed. The first, drawing on Luhmannian systems theory, distinguishes psychic experience, organisational decision structures, and stabilised semantics, treating sociotechnical effects as patterns generated through coupling between technical artefacts and social communication. The second distinguishes meta-, macro-, meso-, and micro-dynamics shaping economic imperatives, regulation, organisational practice, and lived interaction. The central claim is that intelligence-centred evaluation produces category mistakes across these axes, allowing component-level improvements to be misread as system-level responsibility. This mismatch appears in canonical thought experiments—Searle’s Chinese Room, Block’s Chinese Nation, the “stochastic parrot” critique—which turn on cognition while bracketing moral consequence. It also appears in governance: the EU AI Act takes “trustworthy AI” as its guiding horizon, inviting reliability to stand in for responsibility. Drawing on cognitive psychology, affective computing, and political economy, the paper introduces the Artificial Dyadic Reciprocity (ADR) framework to explain how anthropomorphic design produces sub-reflective capture through aligned linguistic, temporal, and affective cues—a meso-level mechanism shaping micro-level experience, unaddressed at macro-regulatory level, sustained by meta-level growth imperatives. On this basis, the paper proposes moral responsiveness as benchmark: a system-level requirement preserving human responsibility by design through answerability, reciprocity, and enforceable redress. It distinguishes ethics washing from regulatory washing and, drawing on the UK’s Age-Appropriate Design Code, argues that responsiveness constraints are implementable. The provocation is simple: societies may tolerate unintelligent systems, but they cannot sustain architectures that dissolve accountability by design.</p>