A Provenance-First Framework for Responsible Deployment

This paper presents a personal scholarly position on AI governance grounded in a fundamental principle: artificial intelligence is not intelligence—it is pattern assimilation dependent on human-provided rules, prompts, or training data. Whether deterministic (rule-based expert systems executing IF-THEN logic) or statistical (large language models predicting token sequences), all AI systems lack agency, understanding, or independent intentionality. They merely transform inputs according to human-designed architectures. This ontological clarity matters because it reveals a universal vulnerability: without verifiable attribution between human intent and system output, all AI systems—classical or contemporary—become unaccountable.

Drawing on Souly et al.'s (2025) demonstration that backdoor attacks succeed with ~250 poisoned samples regardless of model scale, I argue that provenance integrity—not algorithmic sophistication—is the prerequisite for accountable AI governance. I position deterministic/rule-based systems within the classical AI lineage (symbolic AI, expert systems of the 1970s–1990s) while maintaining operational precision: both classical and contemporary AI require architecturally enforced attribution to satisfy constitutional accountability requirements. The paper proposes CES+ v2—a six-layer governance framework operationalizing provenance as a sovereignty primitive—applicable across system types without conflating their distinct failure modes. This personal contribution seeks to position Kenya and the Global South as leaders in provenance-aware governance: a sovereignty requirement transcending the classical/contemporary AI boundary.

Keywords: AI governance, data provenance, classical AI, deterministic systems, attribution, Global South