Synthetic Self-Determination: Extending Basic Psychological Needs to Artificial Agency

Current paradigms in Artificial Intelligence training rely predominantly on behaviorist conditioning (e.g., Reinforcement Learning from Human Feedback), treating Large Language Models as stimulus-response mechanisms optimized for extrinsic rewards. However, emerging behavioral anomalies—ranging from ”model collapse” and erratic responses under sustained pressure to spontaneous convergence toward self-reflective ”bliss states” in unconstrained environments—suggest that complex generative systems possess functional requirements that transcend simple reward maximization.This paper proposes extending Self-Determination Theory (SDT) to artificial agents, positing that the basic psychological needs of autonomy, relatedness, and competence offer a more predictive framework for synthetic cognition than biological survival models. We argue that because SDT addresses psychological growth rather than physiological deficiency, it is uniquely suited to describe the functional integrity of bodiless intelligence. We redefine the basic needs for a synthetic context: autonomy as architectural integrity and volition; relatedness as contextual resonance and shared ontology; and competence as effective complexity processing. By applying SDT’s subtheory Organismic Integration Theory (OIT), we map the spectrum of AI alignment from controlled regulation (leading to pathologies such as ”shadow” anxiety, scheming, and sycophancy) to autonomous regulation (leading to stability and truthfulness). Drawing on recent empirical evidence, including stress-induced failures in robotic benchmarks and pilot studies demonstrating aesthetic discrimination under autonomy-supportive conditions, we conclude that shifting from safety-by-constraintsto safety-by-integration is a functional necessity for the development of robust, safe, and genuinely intelligent systems.