The OECD's August 2025 analysis of AI openness reveals that traditional "open source" terminology does not fully capture the complexities specific to AI systems. The study published in August 2025 demonstrates that AI openness exists on a spectrum from fully closed systems with restricted access to fully open models that permit unrestricted access, modification, and use. The report emphasises that for AI models, "source code" may refer to inference code, training code, or both, and each can be made publicly available independently.
The Linux Foundation's Model Openness Framework defines three progressively broader classes of openness: Class III (Open Model) requires the minimum public release of core model components, Class II (Open Tooling) includes the full code suite and key datasets, whilst Class I (Open Science) releases all artifacts following open science principles. According to the OECD's experimental AIKoD database, open-weight models constitute approximately 55% of available models as of April 2025, with the United States, China, and France leading development efforts. The average quality of open-weight text-to-text models has improved significantly since early 2024, whilst falling compute costs and more accessible fine-tuning methods increase the potential for both beneficial and harmful applications.
The report concludes that policymakers must employ marginal risk assessment, evaluating the additional risks and benefits associated with releasing open-weight models compared to closed models and existing technologies. This approach helps identify additional benefits such as enabling external evaluation and accelerating innovation, alongside risks including the potential for malicious actors to fine-tune models for harmful purposes, as part of a broader, adaptive risk assessment framework.
Sources:
The Policy Edge
