AI Openness | OECD Risks and Opportunities Guide

Understanding AI Openness and Why Terminology Matters

The rapid proliferation of artificial intelligence systems has brought the concept of AI openness to the forefront of global policy discussions. As governments, researchers, and industry leaders grapple with how to regulate and govern increasingly powerful AI models, establishing a shared vocabulary has become essential. The OECD’s August 2025 report, “AI Openness: A Primer for Policymakers,” provides a comprehensive framework for understanding what openness means in the AI context and why precision in terminology directly affects the quality of policy outcomes.

Unlike traditional software where the concept of “open source” is well-defined and governed by established licenses, AI systems present unique challenges. An AI model consists of multiple distinct components — inference code, training code, model weights, training data, and documentation — each of which can be shared or withheld independently. This complexity means that describing an AI system as simply “open” or “closed” fails to capture the nuanced reality of how these systems are built, distributed, and used. For organizations exploring how interactive AI-powered content experiences can transform document engagement, understanding these distinctions is critical for making informed technology choices.

The OECD report was approved by the Global Partnership on Artificial Intelligence (GPAI) in June 2025 and reflects extensive consultation across OECD member states. It builds on earlier discussions by the Working Party on AI Governance and incorporates feedback from national delegations, making it one of the most authoritative guides to AI openness terminology available today.

Why Open-Source AI Is a Misleading Legacy Term

The term “open source” was coined in 1998 as a social contract describing software designed to be publicly available under specific licensing conditions. According to the Open Source Initiative (OSI), open-source software licenses must meet ten key criteria, including free access to source code, permission for derived works, and no restrictions on users or use cases. This framework has powered decades of collaborative software development, from the Linux kernel to Apache web servers.

However, applying this terminology to AI creates fundamental conceptual problems. In traditional software, source code is the primary artifact — the set of human-readable instructions from which executable programs are compiled. In AI systems, the situation is fundamentally different. “Source code” could refer to either the inference code used to run a trained model, the training code used to create it, or both. These can be shared independently. More importantly, the trained model weights — the numerical parameters that encode what the model has learned — are not source code at all. They are the output of a training process applied to data.

The OECD report highlights that multiple competing definitions of “open-source AI” exist today. The OSI released a draft definition leveraging the OECD’s own definition of an AI system, focusing on freedoms to use, study, modify, and share. Meanwhile, the Linux Foundation proposed an alternative definition that explicitly requires sharing information about underlying components. This fragmentation creates confusion for policymakers who need clear, consistent terminology to draft effective regulation.

Several high-profile examples illustrate the problem. Meta’s LLaMA models have been described as “open source” despite containing license restrictions that are inconsistent with the OSI definition. Some developers claim open-source status simply because model weights are available for download, even when their licenses restrict certain use cases and distribution. The OECD therefore recommends using the term “open-weight models” to describe foundation models whose trained weights are publicly available — a more precise term that avoids the conceptual baggage of “open source.”

The Spectrum of AI Openness From Closed to Fully Open

One of the report’s most important contributions is its articulation of AI openness as a continuous spectrum rather than a binary state. At one extreme sit fully closed systems like DeepMind’s Chinchilla, where access to the model and its underlying data is highly restricted. At the other extreme are fully open models like GPT-J, which make all training code, inference code, weights, and documentation publicly available for unrestricted use, modification, and distribution.

Between these poles lies a rich landscape of intermediate options. Cloud-based API access models, exemplified by OpenAI’s GPT-4 and Anthropic’s Claude, allow users to interact with the model through controlled interfaces without downloading any components. Gated downloadable models restrict access to approved users, while non-gated downloadable models provide some components freely while withholding others. Each configuration carries different implications for innovation, security, accountability, and competition.

The Linux Foundation’s Model Openness Framework (MOF) provides a structured approach to evaluating and classifying this spectrum. It identifies 17 critical components for a complete model release — spanning code, data, and documentation — and defines three progressively broader classes of openness. Class III (Open Model) requires releasing the core model architecture and parameters. Class II (Open Tooling) adds the full suite of training and evaluation code plus key datasets. Class I (Open Science) represents the apex, requiring all artifacts including raw training data, research papers, intermediate checkpoints, and log files.

This framework is particularly valuable for policymakers because it provides concrete criteria for evaluating claims of openness. Rather than accepting vague assertions that a model is “open,” regulators can assess which specific components have been released and under what terms. Understanding these gradations helps organizations seeking AI-driven transformation evaluate technology partners and select models that match their transparency and governance requirements.

Licensing Frameworks and Their Impact on AI Innovation

Licensing in open-source AI governs the terms under which models, their weights, and associated code can be used, modified, and distributed. Unlike proprietary models that restrict access to discrete users or customize licensing conditions, open-source AI offers various standardized licensing schemes that grant different nonexclusive conditions to all users, generally at no cost. The choice of license can significantly impact collaboration, innovation speed, adoption rates, and the potential for both beneficial and harmful uses.

Permissive licenses such as Apache 2.0 and MIT allow broad usage, modification, and commercialization with minimal restrictions beyond attribution and disclaimers. These licenses have driven rapid innovation by enabling widespread experimentation and integration into commercial products. The Apache 2.0 license, for instance, has become the default choice for many major open-weight model releases because it strikes a balance between openness and legal clarity.

Conversely, copyleft licenses — like certain versions of Creative Commons — require that derivative works also be shared under the same or similar open-source terms. While these licenses aim to ensure the continued openness of the AI ecosystem, they can create friction for commercial adoption. The OECD notes that more restrictive licenses may be necessary to incentivize model development and investment for specific market needs, but they could also inadvertently limit collaboration and the pace of innovation.

The Linux Foundation’s MOF addresses licensing by stipulating specific license types for different artifact categories: open-source licenses for code (Apache 2.0, MIT), open-data licenses for datasets and model parameters (CDLA-Permissive, CC-BY), and open-content licenses for documentation (CC-BY). This granular approach recognizes that a single license cannot adequately cover all components of an AI system, and that different artifacts require different legal frameworks to balance openness with appropriate protections.

Evolution and Market Trends in Open-Weight Models

Since OpenAI launched ChatGPT in November 2022, the market for generative AI foundation models has experienced extraordinary growth. Using experimental data from the OECD’s AIKoD database — which tracks active generative AI models available as AI-as-a-Service across 51 cloud providers in 11 countries — the report documents a marked acceleration in the global supply of foundation models, particularly from mid-2024 onward.

The data reveals that open-weight models have not merely kept pace with market growth but have become the majority. As of April 2025, open-weight models account for approximately 55% of all commercially available foundation models offered through API endpoints. This represents a significant shift from earlier years when closed models dominated the landscape. The trend suggests that the AI industry is increasingly embracing openness as a development and distribution strategy.

Performance quality has also improved dramatically. Open-weight foundation models have achieved significantly higher scores on common benchmarks since early 2024, narrowing the gap with leading closed models. Models like DeepSeek R1, Meta’s LLaMA 3, and Alibaba’s Qwen have demonstrated that open-weight development can produce state-of-the-art results. This quality improvement, combined with falling compute costs and more accessible fine-tuning methods, has lowered barriers to both use and beneficial applications — while simultaneously increasing the potential for misuse.

The OECD’s AIKoD database provides particularly valuable insights because it tracks not just model availability but also pricing, quality benchmarks, and the relationship between model developers and cloud providers. As of May 2025, it includes models from developers in 14 countries, offered by providers in 11 countries, offering a uniquely comprehensive view of the global AI model ecosystem. For professionals tracking how AI transforms business document workflows, these market dynamics directly influence the tools and capabilities available.

Global Distribution of Open-Weight AI Development

The geographic distribution of open-weight AI development reveals important patterns about the global AI ecosystem. The United States leads decisively in both model development and provision, reflecting its robust AI ecosystem, deep talent pool, and extensive cloud infrastructure. American companies and research institutions have produced the largest number of open-weight foundation models, and US-based cloud providers host the most diverse selection.

China and France emerge as the second and third largest developers of open-weight models, respectively. China’s position reflects massive government and private sector investment in AI development, with companies like Alibaba, Baidu, and DeepSeek producing internationally competitive models. France’s prominence is largely driven by Mistral AI, which has positioned itself as Europe’s leading open-weight model developer and has attracted significant investment to build competitive alternatives to American and Chinese models.

Interestingly, the largest provider hubs do not necessarily align with the largest development centers. The Netherlands and Singapore stand out as major hosting locations despite having fewer domestic model developers. This divergence underscores the global nature of AI deployment, where models are often hosted in countries with advanced cloud capabilities, favorable regulatory environments, and strategic geographic positions — regardless of where they were originally developed. The data reveals a growing international dispersion of model provision, suggesting that the AI services market is becoming more geographically distributed over time.

This geographic analysis has direct policy implications. Policymakers in countries that are primarily consumers rather than developers of AI models need different strategies than those in leading development nations. Countries serving as provider hubs face unique regulatory challenges around liability, data sovereignty, and cross-border AI governance that differ from those faced by model developer nations.

Benefits of Openly Releasing Foundation Model Weights

The OECD report identifies several significant benefits of releasing foundation model weights publicly. First, openness enables external evaluation and accountability. When researchers and auditors can inspect model weights, they can conduct independent assessments of bias, safety, and performance that would be impossible with closed models. This transparency is critical for building public trust in AI systems and for enabling the kind of rigorous scientific evaluation that drives genuine improvement.

Second, open-weight models accelerate research and innovation. By providing a shared foundation that researchers worldwide can build upon, open-weight releases reduce duplicative effort and enable faster iteration. Academic institutions, small companies, and developers in emerging economies gain access to state-of-the-art capabilities that would otherwise be available only to the largest technology companies. This democratization of access has been particularly impactful in specialized domains like healthcare, where researchers at the National Institutes of Health and university hospitals have fine-tuned open-weight models for clinical applications.

Third, open-weight models foster competition by preventing excessive concentration of AI capabilities in a small number of firms. When powerful models are freely available, the competitive advantage shifts from raw model capability to implementation quality, domain expertise, and user experience. This dynamic creates a more level playing field and reduces the risk that a handful of companies will control the trajectory of AI development.

Fourth, local deployment of open-weight models supports sensitive data management. Organizations handling confidential information — healthcare providers, financial institutions, government agencies — can run models on their own infrastructure without sending data to external APIs. This capability is essential for compliance with data protection regulations such as the EU’s GDPR and for maintaining control over proprietary information.

However, the OECD cautions that realizing these benefits depends on access to sufficient computing resources, data, and skilled talent. Open-weight models may be freely available, but the infrastructure needed to run, fine-tune, and deploy them at scale remains expensive and requires specialized expertise. This means that the benefits of openness are not equally distributed — they accrue disproportionately to actors with greater technical and financial resources.

Security Risks and Malicious Use of Open-Weight Models

Alongside the benefits, the report documents significant security risks associated with open-weight models. The availability of model weights empowers malicious actors to fine-tune models for unintended and harmful uses. Unlike closed models where providers can enforce safety policies through API controls, open-weight models can be modified to bypass safety guardrails established by original developers.

Specific risk categories identified in the report include the generation of deepfakes — synthetic media designed to deceive — which can be used for fraud, political manipulation, and reputation attacks. Advanced cyberattacks become more accessible when malicious actors can fine-tune language models to generate sophisticated phishing campaigns, identify software vulnerabilities, or create malware. The report also flags the serious risk of large-scale generation of child sexual abuse material (CSAM) and non-consensual intimate imagery (NCII), areas where open-weight models have lowered the technical barriers to content creation.

The OECD also addresses the more speculative risk of misuse in areas like biology and chemistry, where foundation models could potentially assist bad actors in developing biological or chemical agents. While the marginal contribution of AI models to these risks relative to existing knowledge sources is debated among experts, the report emphasizes that rapid improvements in model quality are increasing potential risks in these domains.

A particularly challenging aspect of open-weight model security is the irreversibility of release. Once model weights are made publicly available, they cannot be recalled or retroactively restricted. Unlike cloud-based APIs where providers can update safety measures or revoke access, publicly released weights exist permanently in the wild. Malicious actors with sufficient expertise and computing resources can fine-tune these models to enhance their propensity for misuse, creating a persistent security challenge that cannot be addressed through post-release interventions.

The report recommends that decisions to release model weights should carefully consider potential benefits and risks, acknowledging that falling compute costs and more accessible fine-tuning methods are lowering barriers to both use and misuse. This dual-use challenge — where the same openness that enables beneficial innovation also facilitates harmful applications — is the central tension that policymakers must navigate.

Marginal Risk Assessment and Policy Recommendations

One of the report’s most sophisticated contributions is its introduction of the concept of marginal risk assessment for open-weight models. Rather than evaluating the risks of open-weight models in isolation, the OECD recommends assessing the marginal or incremental risks and benefits of releasing weights compared to what is already achievable with closed models and existing technologies. This framing is critical because many of the harmful activities attributed to open-weight models — such as phishing, misinformation, or simple cyberattacks — were already possible before these models existed.

The marginal risk framework asks: does releasing model weights meaningfully increase the capability of malicious actors beyond what they can already achieve? For some risk categories, the answer is nuanced. Generating convincing text-based disinformation, for example, was already feasible with closed-source models and even with pre-AI tools. The marginal risk from open-weight models may be limited in such cases. For other categories — such as removing safety guardrails from powerful models, creating specialized harmful content, or enabling entirely new attack vectors — the marginal contribution of open weights may be more significant.

The OECD emphasizes that marginal risk assessment should not be the only tool in the policy toolkit. It should be embedded within a broader, holistic risk assessment framework that can adapt to evolving capabilities and usage patterns. As models become more capable and as fine-tuning becomes easier, the marginal risk calculus may shift. Policymakers should establish monitoring mechanisms that can detect and respond to these changes over time.

The report also highlights the concept of opportunity cost — the potential downsides of not releasing open-weight models. Excessive restrictions on openness could stifle innovation, limit external scrutiny, concentrate power in a small number of companies, and deprive developing countries of access to transformative technology. Balancing these opportunity costs against security risks is the fundamental challenge that the OECD’s framework is designed to address, reflecting the broader tension between innovation and governance that defines AI policy globally.

Conclusion: Balancing AI Openness With Responsible Governance

The OECD’s primer on AI openness provides an essential foundation for the policy debates that will shape the future of AI governance. By clarifying terminology, documenting market trends, and articulating both the benefits and risks of open-weight models, the report equips policymakers with the conceptual tools needed to navigate one of the most consequential decisions in AI regulation: how open should AI be?

The answer, as the report makes clear, is not a simple one. AI openness exists on a spectrum, and the optimal position on that spectrum depends on the specific context — the capabilities of the model, the intended use case, the available safeguards, and the broader technological landscape. What the report provides is a framework for making these decisions thoughtfully, based on evidence and analysis rather than ideology or commercial interest.

As open-weight models continue to improve in quality and grow in market share, the stakes of this debate will only increase. Models that were cutting-edge today will be freely available tomorrow, and the pace of advancement shows no signs of slowing. Policymakers who invest in understanding AI openness now — its definitions, its spectrum, its benefits, and its risks — will be far better positioned to craft governance frameworks that promote innovation while protecting against harm. The OECD’s primer is an indispensable guide for this journey, offering clarity and rigor in a field where both are often in short supply.

For organizations and policymakers seeking to make dense AI research more accessible, transforming reports like this into interactive experiences can dramatically increase engagement and comprehension across diverse stakeholder groups.