Inference as the defining criterion of AI: rethinking the EU AI Ac

This article examines how inference operates as a legal and operational threshold in the EU Artificial Intelligence Act. It argues that a system's capacity to infer how to generate outputs, rather than simply executing predefined instructions, is the key criterion distinguishing AI from traditional automation. Tracing the evolution of the Act's definition of an ‘AI system’, the article shows how inference emerged as its core distinguishing concept, shaping both scope and regulatory purpose. Drawing on computational theory and safety-critical regulation, it proposes a problem-oriented framework for evaluating AI systems in legal contexts. Instead of relying on abstract or exhaustive definitions, this approach classifies AI systems by the problems they solve, the inference methods they employ, and the risks they pose. The framework accommodates hybrid systems, including neuro-symbolic and quantum-enhanced models, offering a more precise, technology-neutral basis for legal interpretation, compliance and enforcement under the AI Act.