Educational AI at the boundary of high risk: Estonia’s AI Leap as a stress test for the EU AI Act and the GDPR

Educational artificial intelligence (AI) does not become high-risk under the European Union Artificial Intelligence Act 2024 merely because it is used in schools. Its classification depends on the function it performs within the educational process: whether it supports learning formatively, or whether it evaluates learning outcomes in ways that affect a student’s educational trajectory. This distinction is especially unstable in adaptive generative AI systems, where the same interactional features that make a tool pedagogically useful may, if repurposed or institutionally relied upon, move it towards the assessment logic of Annex III, point 3(b) of the AI Act. This paper uses Estonia’s AI Leap initiative as an early test case for that boundary. It develops four claims. First, the boundary between formative tutoring and high-risk assessment is conditional and dynamic, rather than inherent in the technology itself. Second, the profiling rule in Article 6(3) of the AI Act is best understood as an anti-derogation trigger: it does not independently bring a system within Annex III, but once an educational AI system falls within Annex III, profiling prevents reliance on the derogation from high-risk classification. Third, the initiative exposes a value-chain asymmetry characteristic of public-sector educational deployments built on third-party general-purpose AI models, in which downstream actors may acquire compliance responsibilities without full technical control. Fourth, Section 20¹ amendment to the Basic Schools and Upper Secondary Schools Act (Estonia) 2010 provides an unusually concrete national-law basis for educational AI processing, but it cannot by itself resolve the data-protection problems created by open-input generative systems, especially the incidental capture of special-category and third-party data. The paper concludes that the initiative’s regulatory significance lies less in its present classification than in the governance conditions needed to prevent functional drift and mitigate the data-protection risks created by open-ended student inputs.