Between legal aspiration and engineering reality: technical and legal barriers to implementing control safeguards for high-risk AI systems

Artificial intelligence (AI) is increasingly used in decisions that affect health, safety, welfare and rights, and regulators have responded by requiring “control safeguards”, such as human oversight, robustness and monitoring, for systems classed as high-risk. This article asks why such safeguards are hard to implement and why the difficulty matters for society. It combines doctrinal analysis of the European Union’s Artificial Intelligence Act (Regulation (EU) 2024/1689, as amended in 2026) with a structured critical review of the technical, human-factors and policy literature. Five technical barriers are identified (model opacity, speed and autonomy, brittle guardrails, limits of human attention and judgement, and the lack of tests for control effectiveness), together with six legal and institutional barriers (open-textured legal standards, uncertain classification, delayed standards and guidance, limited regulatory capacity, fragmented enforcement, and burdens on small organisations). The article argues that these barriers reinforce one another through three gaps: a feasibility gap between what the law demands and what technology can deliver, a verification gap between legal requirements and testable criteria, and a capacity gap between large and small actors and between providers and regulators. Mapping each oversight capability in Article 14(4) to these barriers shows where safeguards are most likely to be formal rather than effective. The article proposes ten recommendations for EU institutions, standard-setters, national authorities, providers and deployers, and considers lessons for lower-capacity jurisdictions, including in Africa, that are drawing on EU-style models. The central conclusion is that regulatory attention should shift from documenting oversight to demonstrating that it works.