The Digital Psyche

Some recurrent AI failures appear as patterns across situations rather than as defects visible in a single output. This Perspective asks when psychology helps explain those patterns without turning metaphor into diagnosis. A pattern must first remain reliable under irrelevant perturbations, predict behavior on an independent task, and change under selective intervention. Those three gates bear on a candidate mechanism. The psychological name faces another test: it must add value through prediction, causal discrimination, or intervention choice beyond what a matched engineering account provides; otherwise, we should keep the engineering result and release the clinical name. The paper applies this rule to automatic perspective sensitivity, aversive learning, and workspace-like integration, where current evidence does not establish a machine disorder or human-like mechanism but does motivate each inquiry across current models and agentic systems. We then turn the psychological lens toward the human–AI interface and develop the Harm-Gated Workspace as a bounded architectural hypothesis. Three comparative protocols make the central claims rejectable. Machine Psychology earns its place through experiments, not labels.