AI Act: researchers propose testing concepts before laws

Testing legal concepts before fixing them in legislation is the proposal advanced by Claudio Novelli, Luciano Floridi, Sofia Ranchordás, Antonino Rotolo, Leendert Van der Torre and Liuwen Yu in “Conceptual Sandboxes: Testing Legal Concepts Before It’s Too Late”, published on Oxford Law Blogs on 10 November 2026. These “conceptual sandboxes” would allow lawmakers and stakeholders to compare alternative definitions and their consequences before codification. The academic proposal addresses the risk that agreement on legislative vocabulary conceals differences over its meaning, leaving authorities, businesses and courts to resolve them.

The authors use the fundamental rights impact assessment under Article 27 of the European Union’s AI Act as an example. In their reading, the regulation’s product safety approach creates tensions with an assessment focused on the context of use and imbalances of power. The commentary particularly questions the clarity of the consequences when an assessment identifies risks that did not emerge during the system’s conformity assessment. This is an interpretative criticism: the account of Article 27 has not been checked directly against the legislative text and does not independently establish the duty’s scope or application.

The proposed protocol has six phases: identify decisive concepts, predict their effects, test predictions through hypothetical cases, compare alternatives, repair or replace categories, and audit outcomes after legislation is adopted. For the “deployer”, the authors suggest differentiated categories based on public functions and the degree of power imbalance relative to the people affected. This is a recommendation for conceptual reform, rather than a classification presented as current law. The model provides for participation and reasoned choices, while acknowledging that making disagreements explicit can complicate political compromise.