Language as a revisable control representation in human and artificial systems

Language can encode conditional rules that are maintained and revised during action. This Perspective proposes relational self-cuing: familiar linguistic constructions help retrieve control rules and bind their variables to current goals, evidence, and actions. The hypothesis predicts a natural-language advantage over informationally matched encodings on novel rule combinations, with a larger advantage there than on practiced mappings. Human task-switching findings motivate this hypothesis, while evidence separating language from domain-general cognition constrains its scope. Global workspace accounts supply a question about selective availability for control, rather than a warrant for attributing consciousness to language agents. The Affective-Cognitive Operating System is reformulated as an unvalidated reference specification for testing the hypothesis. Its proposed implementation distinguishes model-generated records, externally enforced action constraints, and independent execution evidence. A proposed routing experiment would separate causal record use, representational format, and appraisal-dependent policies. A corresponding prospective human experiment would test conditional linguistic self-cuing against general working-memory load. Selective intervention effects and advantages over matched controls would support bounded claims about control; sufficiently precise equivalence would favor simpler explanations. Neither outcome alone establishes phenomenal consciousness.