Artificial Creativity: From Predictive AI to Generative System 3

Juan Carlos Chavez-Autor^1,2,3*

• ¹Universidad Panamericana, Mexico City, Mexico
• ²Goteborgs universitet, Gothenburg, Sweden
• ³Keiser University, Fort Lauderdale, United States

Large language models generate fluent text yet often fail to sustain novelty, task relevance, and diversity across extended contexts. We argue this shortfall persists because current systems implement only fragments of a tri-process loop that supports human creativity: spontaneous ideation in the default-mode network (DMN; broadly System 1–like), goal-directed evaluation in the central-executive network (CEN; broadly System 2–like), and a metacognitive integrator—System 3—that, via neuromodulatory gain control, shifts between exploration and focused control. We introduce Generative System 3 (GS-3), an architecture-agnostic design pattern with three roles: a high-entropy generator, a learned critic, and an adaptive gain controller. Beyond "pure prediction" and simple "reflective prompting," GS-3 identifies the missing pieces for Artificial Creativity: an internal evaluator, endogenous control over sampling entropy, and adaptive priors maintained across extended contexts. This Conceptual Analysis (i) formalizes novelty, usefulness, and diversity with operational definitions; (ii) develops multiple gain-update policies (exponential, linear, logistic) with stability constraints and sensitivity expectations; (iii) derives falsifiable behavioral indices—associative-distance density, analytic-verification ratio, and convergence latency—with pass–fail criteria; and (iv) provides a proof-of-concept blueprint and evaluation protocol (tasks, metrics, ablations, reproducibility kit). We position GS-3 relative to computational-creativity and co-creative frameworks, and delineate where brain–model analogies are functional rather than literal. Ethical guidance addresses bias, cultural homogenization, and reward gaming of proxy objectives (often termed "dopamine hacking") through plural critics, transparent logging, and outcome-tied entropy caps. The result is a testable roadmap for transitioning from regulated prediction to genuinely creative generative systems.