Article language: English.
Abstract
The EU Artificial Intelligence Act enables Member States to jointly create AI regulatory sandboxes and guarantees mutual recognition of participation in AI sandboxes across the Union. But the Act does not fully describe how such joint sandboxes should work when testing, supervision, evidence generation and potential harm is beyond the confines of a single national legal order. This paper discusses whether a cross-border AI sandbox, with a possible Spain-Czech partnership in the middle being the central focus, is both legally possible and practically viable in the light of Articles 57 and 58 of the AI Act. It starts with an explanation of the legal nature of AI sandboxes, alongside an argument that the Act establishes a guidance-based and supervision-centered model, as opposed to a broad derogation regime. It then examines four operational gaps that may prevent mutual recognition becoming effective in practice: the narrowness of the capacity of the sandbox to overcome hard law limits; the lack of common admission criteria for joint participation; institutional differences in the distribution of supervisory authority; and the need to have an enforcement backstop which can give cross-border effect to supervisory interventions and sandbox outputs. The remaining problem in private law, which the article discusses, is the AI Act preserves the liability under the ordinary Union and national law, but leaves the question of which court has the competence to hear the case. The article concludes that cross-border sandboxes can exist, albeit with support of a detailed bilateral or multilateral protocol containing admission criteria, supervisory roles, escalation procedures, interoperable records, financial assurance and evidence-preservation duties.