The Emergence of Artificial Intelligence on the Policy Agenda: Insights from the 2025 OECD Report

Introduction

In August 2025, the Organisation for Economic Co-operation and Development (OECD) issued a report entitled “AI Openness: A Guide for Policy Makers.” This report provides a comprehensive overview of the concept of openness in artificial intelligence, distinguishing it from traditional software. It analyzes the benefits and risks associated with releasing certain components of AI models openly—specifically open-weight foundational models—with the aim of guiding policy discussions and promoting responsible governance.

Key Concepts and Terminology

The report clarifies that the term “open source,” which originated in the software world, is misleading when applied to artificial intelligence. Unlike a simple block of code, an AI system consists of multiple components—such as training code, inference code, training data, and model weights—each of which can be made publicly available independently. Therefore, AI openness exists on a broad spectrum, ranging from fully closed systems to models in which all components are publicly accessible. The report focuses on open-weight models—foundational models whose trained weights can be downloaded by the public—as they represent the core of current policy discussions.

Benefits and Risks

The report presents a dual perspective on open-weight models:

Benefits

• Accelerating Innovation: Openness enhances research and rapid development by allowing a broad community of developers to build upon and modify existing models.
• Enhancing Competition: By lowering the entry barrier, open-weight models challenge the dominance of a few major companies, fostering a more diverse and competitive AI ecosystem.
• Promoting Transparency and Accountability: Public access to model weights enables external auditing and evaluation, helping to identify and address potential biases, safety issues, and other vulnerabilities.
• Data Privacy: Organizations can use open-weight models on-premise, allowing them to retain sensitive data within their own infrastructure.

المخاطر

• Malicious Use: The availability of these models can enable malicious actors to misuse them for creating harmful content, such as deep fakes, conducting sophisticated cyberattacks, or producing other illegal materials.
• Bypassing Safeguards: Once model weights become public, users can fine-tune them, potentially bypassing the safety filters and controls implemented by the original developers.
• Practical Irrevocability: Once an open-weight model is released, it is nearly impossible to withdraw it or impose new restrictions retroactively, making it a persistent source for potential undesired uses.

Policy Implications

The report concludes by urging policymakers and developers to weigh the marginal benefits and risks of releasing open-weight models. It notes that lower computing costs and simplified fine-tuning methods have reduced barriers for both beneficial and malicious use. The decision to release a model should be part of a comprehensive risk assessment that considers whether the benefits outweigh the potential harms in a given context, while also acknowledging the opportunity cost of not fostering a more open AI ecosystem.