Scholarlink

<p>This paper critically examines the EU Artificial Intelligence Act’s definition of “Free and Open Source AI” (FOS AI), arguing that its current approach risks endorsing superficial openness while neglecting both practical and ideological ramifications: on the one hand, the ability to inspect and improve a system, thereby building trust and furthering innovation; on the other, the substantive principles of freedom and transparency foundational to the Free and Open Source Software (FOSS) movement. Drawing on the Four Freedoms of FOSS including the rights to use, study, modify and distribute, we analyze what it genuinely means for an AI system to be “free” and identify the technical and legal components required to meet this standard. Our analysis reveals a critical flaw in the Act’s approach: by permitting partial disclosures (e.g., model weights without training data) and descriptive dataset summaries in lieu of full transparency, the framework inadvertently enables “open-washing’’. This loophole not only erodes transparency but also creates numerous challenges including an accountability vacuum, where models distributed through informal or extraterritorial channels can bypass the Act’s “provider” and “deployer” categories, leaving downstream users exposed to opaque risks without any clearly identifiable entity responsible for compliance or remediation. To address these challenges, we argue that the AI Act should heighten the threshold for granting exceptions to FOS AI by mandating the disclosure of all relevant AI components, including substantial access to training datasets. We acknowledge that dataset transparency introduces substantial challenges related to privacy, fairness, and intellectual property, and that it should not be hastily applied to circumstances exceeding the current scope of the EU AI Act, namely non-high-risk products or services where models without presumed systemic risks are utilized. We discuss the desirable outcomes that a stronger threshold for a FOS AI exception would enable towards the stated goals of the EU legislator.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Regulating transparency: a critique of the “free and open source AI” definition in the EU AI act

  • 期刊论文

  • 发表日期：

    2026年5月23日

查看全文
  • Simona Ramos,
  • Fabio Pianese

摘要

This paper critically examines the EU Artificial Intelligence Act’s definition of “Free and Open Source AI” (FOS AI), arguing that its current approach risks endorsing superficial openness while neglecting both practical and ideological ramifications: on the one hand, the ability to inspect and improve a system, thereby building trust and furthering innovation; on the other, the substantive principles of freedom and transparency foundational to the Free and Open Source Software (FOSS) movement. Drawing on the Four Freedoms of FOSS including the rights to use, study, modify and distribute, we analyze what it genuinely means for an AI system to be “free” and identify the technical and legal components required to meet this standard. Our analysis reveals a critical flaw in the Act’s approach: by permitting partial disclosures (e.g., model weights without training data) and descriptive dataset summaries in lieu of full transparency, the framework inadvertently enables “open-washing’’. This loophole not only erodes transparency but also creates numerous challenges including an accountability vacuum, where models distributed through informal or extraterritorial channels can bypass the Act’s “provider” and “deployer” categories, leaving downstream users exposed to opaque risks without any clearly identifiable entity responsible for compliance or remediation. To address these challenges, we argue that the AI Act should heighten the threshold for granting exceptions to FOS AI by mandating the disclosure of all relevant AI components, including substantial access to training datasets. We acknowledge that dataset transparency introduces substantial challenges related to privacy, fairness, and intellectual property, and that it should not be hastily applied to circumstances exceeding the current scope of the EU AI Act, namely non-high-risk products or services where models without presumed systemic risks are utilized. We discuss the desirable outcomes that a stronger threshold for a FOS AI exception would enable towards the stated goals of the EU legislator.