EU AI Act's research carve-outs risk upending how scientists publish findings

The EU AI Act, finalized in 2024, marks one of the world's strictest frameworks for governing artificial intelligence systems—but a forthcoming analysis suggests its research exemptions may have an unintended consequence: making it harder for scientists to publish and share their work openly.

A position paper submitted to arXiv argues that while the regulation carves out certain protections for research activities, the way those exemptions are written could paradoxically disqualify research from protection the moment findings are released publicly, including through peer-reviewed journals and model-sharing platforms.

The Compliance Puzzle for Researchers#

The EU AI Act imposes a raft of documentation, testing, and risk-management obligations on developers of AI systems—requirements originally designed for commercial deployment. The paper's authors contend that many academic teams may not realize these obligations apply to their work. They outline a roadmap for determining whether a given research project triggers the regulation's requirements, walking through everyday research scenarios where compliance needs could arise.

The analysis identifies a critical tension: the Act defines research exemptions narrowly, and those exemptions may vanish once research enters the scientific publication pipeline. This means researchers could face a catch-22—comply with the Act's stringent requirements or forgo public sharing of results.

Gaps in the Current Framework#

The authors flag three main problems with how the Act handles research:

• Narrow exemption scope: The research exceptions don't adequately account for how modern AI science actually works, particularly the emphasis on releasing models, datasets, and systems alongside papers.
• Publication as a compliance breach: Sharing findings through conferences, journals, or platforms like Hugging Face could technically void the exemptions, since the Act's definitions may treat public disclosure as incompatible with research status.
• Legal uncertainty: AI researchers lack clear guidance on whether their projects qualify for protection, creating compliance risk even for good-faith efforts.

Proposed Solutions#

The paper calls for amendments to the Act that would provide legal clarity for the research community. It also offers two interim recommendations for researchers: first, conduct a careful evaluation of their project's regulatory scope before publication, and second, document compliance efforts proactively to demonstrate good-faith attempts to meet obligations.