From Compliance to Understanding: Translating AI Regulation as an Epistemic Integration Problem

The contemporary landscape of Artificial Intelligence (AI) governance is marked by profound epistemic fragmentation. As AI systems continue to develop rapidly, the socio-technical mechanisms designed to regulate them remain divided across legal, technical, and philosophical domains. This fragmentation is increasingly visible in the early implementation and ongoing interpretation of the European Union AI Act. While the Act establishes binding requirements for transparency, risk management, and human oversight, its operationalisation is constrained by a modern form of the “Two Cultures” divide between normative reasoning and technical optimisation.

This paper argues that translating AI regulation into practice is not merely a technical or legislative task, but a challenge of epistemic integration. Drawing on the distinction between performance and understanding, it contends that compliance cannot be reduced to documentation, logging, or statistical traceability. Such practices often produce visibility without providing genuine intelligibility, reinforcing algorithmic opacity under the guise of transparency. For example, documentation artifacts such as model cards provide information about system provenance and performance, but do not ensure that system behaviour becomes understandable to human operators. Similarly, conformity assessments risk becoming procedural checklists that satisfy formal requirements while leaving the question of understanding unresolved.

To address this issue, The paper argues that implementing regulatory compliance requires model-based epistemic reasoning. The EU AI Act provides normative models describing how AI systems should function within society, and these models must be translated into meaningful computational and organisational practices through interpretation, judgement, and interdisciplinary coordination. The paper proposes a framework of epistemic integration in which philosophy serves as a mediating practice that aligns normative, technical, and organisational forms of knowledge. Concretely, this framework maps obligations such as human oversight, risk classification, and transparency onto stages of the machine learning lifecycle, from data governance and model training to deployment and post-market monitoring. It also identifies where purely technical compliance measures fail to deliver genuine intelligibility.

The paper concludes that effective AI governance requires epistemic integration: the deliberate alignment of knowledge across fragmented communities as a condition for preserving meaningful human control and the safe operation of complex socio-technical systems. Philosophy is not an optional supplement to this process. It is the practice through which regulatory frameworks become intelligible, actionable, and practically effective, providing a necessary foundation for governing AI systems whose consequences extend far beyond the reach of any single discipline.