What Open Source Means
The term "open source" has a specific definition. In software, the Open Source Initiative has maintained the Open Source Definition since 1998. It requires free redistribution, source code access, the ability to create derived works, and no discrimination against persons, groups, or fields of endeavor. The definition was created to provide clarity about what "open" means in practice — to prevent the term from being used to describe software that is available to inspect but not to modify, or free to use but not to distribute.
In October 2024, the OSI extended this definitional work to AI with the Open Source AI Definition (OSAID). The OSAID requires that an open-source AI system provide access to the complete code used for training, inference, and evaluation; comprehensive information about training data sufficient to reproduce the system; and the model parameters (weights) in a format that allows study, modification, and redistribution. The definition was developed through a multi-stakeholder process involving industry, academic, and civil society participants.
The OSAID matters for AI governance because the EU AI Act includes exemptions for open-source AI models. If a model qualifies as open source, certain regulatory obligations are reduced. The definition of what counts as "open source" is therefore not merely a labeling question — it is a regulatory question with direct consequences for which models are subject to which obligations.
Meta's Llama and the Definition Gap
Meta releases its Llama models under terms that Meta describes as "open source." Llama model weights are publicly available. The code for inference is publicly available. Meta publishes research papers describing the model architecture. By the standards of pre-OSAID AI practice, Llama is among the most open frontier model releases.
By the standards of the OSAID, Llama does not qualify as open source. The Llama license includes commercial restrictions for applications with more than 700 million monthly active users — a restriction that is incompatible with the open-source requirement of no discrimination against fields of endeavor. Meta does not provide access to the training data or comprehensive documentation of the data used to train Llama models, making the system impossible to reproduce independently. The model weights are available, but the complete system is not.
Meta and OSI publicly disagreed in October 2024 over the definition. Meta maintained that its Llama release strategy constitutes open source. The OSI's published definition says otherwise. Meta is currently lobbying EU regulators to accept Meta's version of "open source" for purposes of the AI Act's exemptions — which would allow models that do not meet the OSI definition to claim the regulatory benefits of open-source status.
The distinction matters precisely because Meta has deployed the "open source" label as part of a regulatory strategy. If the EU AI Act's open-source exemptions apply to models that meet the OSI definition, Llama would be subject to the full regulatory framework. If the exemptions apply to models that meet Meta's definition, Llama would receive reduced regulatory obligations. The definition debate is a governance debate conducted in the vocabulary of openness.
Mistral and the EU AI Act
Mistral AI, the French AI company, provides a second case study in the strategic deployment of openness rhetoric. Mistral has released several models under terms that are more open than most competitors but that, like Llama, include commercial restrictions that fall short of the OSI definition. Mistral's license bars the use of certain models and outputs for commercial ventures without a commercial license.
During the EU AI Act negotiations, Mistral lobbied against provisions that would impose obligations on general-purpose AI model providers. The company argued that regulatory requirements would disproportionately burden open-model releases and could deter companies from making models available. The argument positions openness — the release of model weights — as a public good that regulation threatens.
The argument has genuine force. There are legitimate safety and democratization benefits to making model weights available. Independent researchers can study models they could not afford to train. Downstream developers can build applications without depending on API access controlled by a single company. Open weight releases do distribute capability and reduce concentration. These benefits are real.
The structural question is whether the benefits of open weight releases should translate into reduced regulatory obligations for the companies making them. The argument that openness should exempt models from safety requirements assumes that openness and safety are aligned — that making a model available for external inspection necessarily makes it safer. This assumption is contested. Open weight releases also enable fine-tuning that removes safety guardrails, deployment in contexts the original developer did not anticipate, and use by actors whose intentions the developer cannot evaluate.
The Three Functions of Openness
The strategic ambiguity of "open source" in AI governance operates because the term serves three distinct functions simultaneously, and the conflation of these functions is useful to the companies deploying the term.
Genuine benefit: external researchers can study model behavior, identify vulnerabilities, and publish findings. Open access enables independent safety evaluation. This function is real and valuable.
Meta's open release of Llama creates an ecosystem of developers building on Meta's infrastructure, competes with OpenAI's closed API model, and positions Meta as the platform layer. Openness is a market-making strategy.
Open-source framing argues against regulation: "You can't regulate open models the same way you regulate closed ones." The EU AI Act's open-source exemption is the specific governance consequence of this argument.
The three functions are not mutually exclusive. A model release can simultaneously provide genuine safety benefits, serve competitive strategy, and function as a regulatory argument. The structural problem is that the third function — the regulatory shield — exploits the legitimacy of the first function — genuine safety benefit — to serve the second function — competitive advantage through reduced regulatory burden. When Meta argues that regulating Llama would harm the open-source ecosystem, the argument borrows the legitimacy of genuine open-source safety benefits to advance a competitive strategy that depends on regulatory exemption.
The Open-Washing Problem
A 2024 paper presented at the ACM FAccT conference — "Rethinking Open Source Generative AI: Open-Washing and the EU AI Act" — documented the phenomenon of "open-washing" in AI: the use of open-source terminology to describe releases that do not meet open-source criteria. The paper examined the EU AI Act's open-source provisions and found that the Act's definitions are ambiguous enough to allow models that are not genuinely open source to claim open-source exemptions.
The open-washing problem is not merely a labeling dispute. It has direct governance consequences. If models that restrict commercial use, that do not provide training data access, and that cannot be reproduced independently can claim open-source status for regulatory purposes, then the open-source exemption functions not as a recognition of genuine openness but as a loophole that rewards the strategic use of openness language.
The term "open source" in AI governance has become a contested definitional space where the regulatory consequences of the definition are larger than the technical consequences.
The definitional contest is not abstract. Meta is spending resources lobbying EU regulators on the definition. Mistral's lobbying against AI Act provisions used open-source arguments. The companies deploying openness rhetoric are the companies with the most to gain from the regulatory consequences of the definition they advocate. This is not conspiracy. It is the documented intersection of corporate strategy and regulatory design in a domain where the technical vocabulary has governance implications.
The Genuine Tension
The honest version of the open-source AI debate contains a genuine tension that the strategic deployment of openness rhetoric obscures. Open weight releases do provide democratization benefits. They do enable independent research. They do reduce the concentration of AI capability in a handful of closed-API providers. These benefits are real and the people who advocate for them — including many researchers and developers with no corporate lobbying interest — advocate in good faith.
Open weight releases also enable proliferation. A model released with open weights can be fine-tuned to remove safety guardrails. It can be deployed by actors who would not pass a responsible-use screening. It can be used in contexts that the releasing company did not anticipate and cannot monitor. These risks are also real, and the people who raise them — including safety researchers within the companies making open releases — raise them in good faith.
The genuine tension is between democratization and proliferation, between transparency and control, between the benefits of open access and the risks of open access. This tension cannot be resolved by the strategic deployment of the "open source" label. It requires governance frameworks that distinguish between genuine openness and strategic openness, that provide appropriate regulatory treatment for each, and that do not allow the vocabulary of openness to function as a blanket exemption from safety obligations.
The Openness Inversion — Named
The strategic deployment of openness language to achieve outcomes that are structurally opposed to the transparency and accountability that genuine openness provides. The Openness Inversion operates when "open source" is used not to describe a model that meets established open-source criteria but to claim regulatory exemptions, to position against closed competitors, and to argue against safety requirements — all while the model in question restricts commercial use, withholds training data, and cannot be independently reproduced. The inversion is complete when the vocabulary of transparency is deployed to reduce rather than increase accountability: when calling a model "open" makes it subject to fewer regulatory obligations than a closed model, despite the open model providing less than full transparency about its training, data, and capabilities.
The Openness Inversion is the fourth element in the governance capture pattern. The Governance Lag creates the vacuum. The Expertise Capture fills it with industry-dependent governance bodies. The Voluntary Commitment substitutes pledges for binding regulation. The Openness Inversion deploys the vocabulary of transparency to argue against the accountability that transparency should enable. The final paper in this series synthesizes these four elements into the structural pattern that connects AI governance capture to the identical pattern documented across regulated industries.
What Genuine Openness Would Require
If openness is to serve its claimed safety function — enabling independent evaluation, external accountability, and transparent governance — then genuine openness in AI requires more than the release of model weights under restrictive licenses. It requires training data documentation sufficient for independent reproduction. It requires capability evaluations conducted by independent parties with the resources to conduct them. It requires safety testing results published in formats that enable external verification. And it requires that the regulatory treatment of open models be conditioned on meeting these criteria, not on self-declaration.
The OSI's OSAID represents an attempt to establish such criteria. The resistance to that definition from companies that use the "open source" label demonstrates precisely why the definition matters: the companies most invested in the label are not the companies most invested in meeting its requirements. The label serves a function that is distinct from — and in some cases opposed to — the substance it describes.
The governance implication is that "open source" in AI policy should be treated as a specific, defined category with specific requirements, not as a self-declared label that triggers regulatory exemptions. What Meta calls Llama and what the OSI calls open source are different things. Governance frameworks that treat them as the same thing are not recognizing openness. They are rewarding the strategic deployment of openness vocabulary.