The EU AI Act requires every member state to establish at least one AI regulatory sandbox by August 2026. These controlled environments — where AI systems can be tested under regulatory supervision before full deployment — represent the Act's most ambitious attempt to reconcile innovation promotion with safety assurance. But operationalizing sandboxes at scale raises challenges that the legislation acknowledges but does not resolve.
Operationalizing Sandboxes
Ahern (2025) provides the most detailed analysis of what operationalizing AI regulatory sandboxes under the EU AI Act actually requires. Sandboxes must balance three competing demands: providing genuine regulatory flexibility (so innovators have reason to participate), maintaining meaningful oversight (so the sandbox is not merely a deregulation zone), and generating transferable knowledge (so lessons from sandbox testing inform broader regulatory practice).
The institutional requirements are substantial. Each sandbox needs dedicated staff with both technical and legal expertise, clear entry and exit criteria, monitoring protocols, and mechanisms for translating sandbox outcomes into regulatory guidance. For many member states — particularly smaller ones with limited regulatory capacity — standing up a functioning AI sandbox alongside all other AI Act implementation requirements strains available resources.
The Liability Question
Tran (2025), presenting at AIES, raises a question that sandbox designs typically avoid: who bears liability when AI systems tested in sandboxes cause harm? Current product liability frameworks may not apply within sandboxes, where systems are explicitly deployed in experimental conditions. But the harm is real — a medical AI tested in a sandbox that misdiagnoses a patient causes genuine injury, regardless of the regulatory status of the test.
Tran proposes no-fault compensation mechanisms for sandbox participants — a form of insurance that separates the question of harm redress from the question of fault attribution. This approach would encourage sandbox participation by reducing liability risk for innovators while ensuring that affected parties are compensated. The proposal draws on precedents from clinical trial regulation, where participants in experimental medical treatments are protected by compensation frameworks that do not require proving negligence.
The Balance Point
Wang (2025) examines the broader question of how sandboxes balance innovation incentives with safety requirements. The evidence from existing sandbox programs — in fintech, data protection, and energy — suggests that the balance depends critically on sandbox governance design. Sandboxes with weak oversight attract participants seeking to avoid regulation rather than engage with it constructively. Sandboxes with excessive oversight deter participation and produce no useful insights. The productive middle ground involves structured flexibility — clear rules about what can be tested, transparent reporting requirements, and genuine regulatory engagement throughout the testing period.
The sandbox experiment is ultimately a test of whether regulation and innovation can be genuinely complementary rather than inherently adversarial. If AI sandboxes succeed, they will demonstrate that structured regulatory engagement can improve innovation quality while building the evidence base for better regulation. If they fail — becoming either rubber-stamp approval zones or bureaucratic obstacles — the adversarial framing will be reinforced, and the implementation gap will widen.
The Knowledge Transfer Problem
Beyond their immediate function of testing specific AI systems, sandboxes are intended to generate knowledge that informs broader regulatory practice. But knowledge transfer from sandbox experiments to general regulatory guidance is not automatic. It requires deliberate documentation, analysis, and dissemination — activities that are often underfunded relative to the sandboxes themselves.
The most valuable sandbox outputs are not approvals or rejections of specific AI systems but insights about how different types of AI systems interact with regulatory requirements in practice. A sandbox that tests a medical AI diagnostic tool generates knowledge about how AI performance should be evaluated in clinical settings, what kinds of evidence satisfy regulatory requirements, and what post-deployment monitoring is appropriate. This knowledge, properly documented and shared, is more valuable than the approval or rejection of any individual system.
Cross-border knowledge sharing between national sandboxes is particularly important given the EU's harmonization goals. If each member state's sandbox develops its own testing methodologies and evaluation criteria independently, the result will be 27 different approaches to the same problems — regulatory fragmentation within the sandbox system designed to reduce it. The European AI Office's coordination role is essential for ensuring that sandbox experiences are shared, methodologies are aligned, and lessons learned in one jurisdiction benefit regulatory practice across the Union.
The participation incentive problem deserves attention. Why would an AI company voluntarily submit its product to regulatory testing when it could simply deploy without sandbox participation? The answer depends on what sandboxes offer: regulatory certainty (a sandbox approval provides stronger legal footing than untested deployment), market access (some procurement processes may prefer sandbox-tested products), and reputational advantage (sandbox participation signals commitment to responsible development). If these incentives are insufficient, sandbox participation will be limited to companies that face regulatory uncertainty anyway, and the sandbox will test only the products that most need testing rather than providing the broad evidence base that effective regulation requires.