A position paper accepted as a spotlight at ICML 2026 warns that the EU AI Act’s research exemptions may be incompatible with the standard publication practices of major AI conferences, creating legal uncertainty for academic researchers who release models and datasets.
The paper, by Alina Wernick and Kristof Meding, argues that the AI Act’s obligations could apply to far more academic AI research than the community currently recognizes. The core problem, they say, is that publishing a model on platforms such as GitHub or Hugging Face may constitute “placing on the market” under the Act’s definition, an action that voids the research exemption.
The exemption paradox
The AI Act exempts AI systems developed and used solely for scientific research. However, the exemption requires that the system is not placed on the market or put into service. The paper argues that the standard practice of AI research, training a model, publishing it with a research paper, and releasing weights or code, can trigger the Act’s compliance obligations, including documentation requirements, risk assessments, and conformity assessments.
This creates what the authors describe as an unintended conflict between the EU’s regulatory framework and the academic convention of open publication. An AI researcher at a European university who trains a foundation model and releases it under an open licence could face the same regulatory obligations as a commercial provider.
What the paper recommends
The authors offer a step-by-step roadmap for researchers to determine whether they need to comply with the Act, along with a visual tool for navigating the complex system of research exceptions. They also propose two practical recommendations: conduct a formal compliance assessment before any public release, and engage with legal experts or institutional review boards early in the research process.
For policymakers, the paper recommends amendments to the AI Act to provide greater legal certainty for academic researchers, specifically clarifying when model and dataset releases trigger regulatory obligations.
Why it matters
The paper highlights a growing tension between the EU’s approach to AI governance and the norms of open AI research. If publishing a model on GitHub is legally equivalent to commercial distribution, European AI researchers face a choice between regulatory compliance and the open publication practices that have driven progress in the field.
As the EU AI Act takes effect in phases through 2026 and 2027, the question of how its rules apply to academic research will become increasingly urgent for institutions, funding bodies, and conference organisers.
Sources: ArXiv (revised July 7, 2026, accepted at ICML 2026)