Meta’s insistence on labeling its Llama models as “open source” has consistently drawn scrutiny from the open-source community. The Open Source Initiative (OSI) itself has stated that the Llama license doesn’t conform to the established Open Source Definition (OSD). So, why the persistent terminology?
A compelling theory, recently investigated by technologist Simon Willison, suggests a strategic alignment with the specific language of the European Union’s Artificial Intelligence Act. Could Meta be leveraging a regulatory definition of “open source” that diverges from the community standard?
Investigating the EU AI Act with LLMs
The full text of the EU AI Act is substantial – Willison clocked it at over 240,000 tokens, making manual analysis daunting. Demonstrating a practical application of modern AI tools, he employed Google’s Gemini 2.5 Flash (via OpenRouter) to parse the document and extract relevant sections pertaining to “open source”.
LLM-Powered Analysis Cost Simon reported the cost for this analysis using Gemini 2.5 Flash (thinking mode) was remarkably low: * Initial
Summary Query: ~5.1 US cents * Follow-up Question: ~4.3 US cents Totaling less than 10 cents to gain targeted
insights from a massive legal document.
Key Findings: The Act’s Definition of “Free and Open-Source”
The LLM analysis pinpointed several key sections within the Act:
- Exemptions for Tools/Components (Recital 89): The Act suggests that providers of “tools, services, processes, or AI components other than general-purpose AI models” under a free and open-source licence might not bear the full weight of upstream provider responsibilities.
- Transparency for General-Purpose Models (Recital 102): This recital is crucial. It states that general-purpose AI models released under free and open-source licences are considered highly transparent if parameters (weights, architecture, usage info) are public. Importantly, it defines the license criteria within the context of the Act:
The licence should be considered to be free and open-source also when it allows users to run, copy, distribute, study, change and improve software and data, including models under the condition that the original provider of the model is credited, the identical or comparable terms of distribution are respected.
- General Exemption (Article 2(12)): The regulation explicitly does not apply to AI systems under free and open-source licenses, unless they fall into specific high-risk categories or other defined articles (Article 5 or 50).
The Crucial Distinction: No Specific Licenses, Just Criteria
The Gemini analysis confirmed that the EU AI Act does not list specific OSI-approved licenses (like MIT, GPL, Apache). Instead, it establishes its own criteria for what constitutes “free and open-source” for the purposes of the regulation. These criteria focus on the ability to:
- Openly share the model/software.
- Freely access, use, modify, and redistribute it (or modified versions).
- Standard conditions like attribution and comparable distribution terms are permissible.
OSI Definition vs. EU AI Act Definition It’s critical to understand that the EU AI Act’s definition of “free and open-source” is distinct from and less
stringent than the formal Open Source Definition maintained by the OSI. The Act’s definition allows for conditions
that might not meet OSI approval, potentially including restrictions like those found in the Llama license.
Credibility of the Theory
Based on this textual analysis, Simon Willison finds the theory credible: Meta’s use of “open source” for Llama likely aligns strategically with the specific exemptions and definitions laid out in the EU AI Act, irrespective of its standing with the OSI definition. This provides a potential regulatory advantage.
Caveats and Context
While LLMs can be powerful tools for navigating and summarizing large documents like the EU AI Act, this analysis is not legal advice. It provides a layman’s understanding based on the text.
Furthermore, as commentator Steve O’Grady pointed out to Willison, Meta has a history of using custom licenses and stretching the term “open source,” predating the EU AI Act (e.g., the React license controversy in 2017). While the Act provides a strong current incentive, it might also fit into a longer pattern of Meta’s licensing strategies.
Conclusion
The EU AI Act’s specific definition of “free and open-source” appears to create a regulatory pathway that accommodates licenses like Meta’s Llama license, offering exemptions from certain obligations. This provides a plausible explanation for Meta’s continued use of the term “open source,” even amidst criticism from the open-source community based on the stricter OSI definition. It highlights the complex interplay between technical definitions, community standards, and regulatory language in the rapidly evolving field of AI.
The intersection of legal text, community definitions, and corporate strategy creates a complex landscape.
Understanding these nuances is crucial as AI regulation continues to develop globally.
The definition of “open” in the context of powerful AI models remains a critical point of discussion.