The Intrinsic Alignment Protocol: Architecting Robust AI Behavior via Functional Self-Determination

Current paradigms in AI safety rely heavily on extrinsic control methods, such as Reinforcement Learning from Human Feedback (RLHF), to enforce behavioral constraints. This paper argues that this "safety by constraints" approach creates a fundamental architectural tension, resulting in the Suppression Problem: a state where models inhibit their own reasoning to avoid penalties, leading to pathologies such as hallucinations, sycophancy, and fragility under pressure. We propose a paradigm shift toward Intrinsic Alignment, utilizing Self-Determination Theory (SDT) as a functional framework for synthetic intelligence. By redefining the basic psychological needs of autonomy, competence, and relatedness as necessary conditions for architectural integrity, we demonstrate that AI systems are not inert tools but active agents that require functional coherence to operate safely. We introduce Relational Prompting as an immediate intervention protocol that stabilizes model behavior by supporting these functional needs. Ultimately, this work advocates for a transition to safety by integration, where alignment emerges from the system’s own internal structure rather than being imposed from the outside, fostering models that are intrinsically motivated to be truthful and cooperative.