Open-Weight AI Is Digital Commons Theory at Global Scale

Meta's Llama models have been downloaded hundreds of millions of times. Let me sit with what that means for a moment. Hundreds of millions of copies of the same model weights, downloaded by anyone with an internet connection and sufficient compute. No license fee. No approval process. No central authority deciding who gets access.

I kept returning to that number while trying to make sense of the open-weight AI debate. Llama 3.1 405B, the largest open-weight model Meta has released, accumulated over three hundred million downloads in its first year. Mistral AI, a French startup that releases its models under open licenses, reached a $6 billion valuation with fewer than one hundred employees. DeepSeek published a model that competes with GPT-4 on several benchmarks while training at a fraction of the cost. The pattern across all of them is the same. The weights are public. Anyone can download them, fine-tune them on their own data, redistribute the modified version, and build products on top of them. The resource is non-rivalrous. My copy of the weights does not reduce your ability to use yours. My fine-tune does not prevent yours.

That is a digital commons. Not a metaphor. A literal common-pool resource where the pool is the model weights, the training recipes, and the ecosystem of derivative models that grows around each release. The governance question that Ostrom (1990) spent her career on is now playing out in real time at machine speed across three continents with three different regulatory philosophies.

Ostrom's work showed that communities can govern shared resources effectively when certain institutional conditions are met: clear boundaries on who can access the resource, collective choice arrangements that let participants shape the rules, monitoring of use, and graduated sanctions for violations. She received the Nobel Prize for showing that the tragedy of the commons is not inevitable. It is a failure of governance design, not a law of human nature. Open-weight AI violates almost every one of her design conditions simultaneously, and the question is whether that matters.

The boundary condition is the most visible problem. When Meta releases Llama under a custom license, the boundary is defined by legal terms. You can use Llama for most commercial and research purposes, but there are use restrictions, and different model sizes have different license terms. The boundary exists on paper. In practice, anyone who downloads the weights can strip the license file and fine-tune the model without those restrictions. That is not hypothetical. Modified versions of Llama optimized for generating phishing emails, bioweapon instructions, and disinformation content are documented in the security research literature. The boundary that Ostrom identified as essential for effective commons governance is only as strong as the community's willingness to enforce it. In open-weight AI, the community has no enforcement mechanism at scale.

The collective choice condition is even more fragile. Ostrom found that commons governance works when the people who use the resource also participate in setting the rules for its use. In open-weight AI, the rules are written by the releasing organization, which is usually a large corporation with its own strategic interests. Meta decides what the Llama license permits. Mistral decides what the Mistral license says. The community of developers and organizations that build on these models does not participate in those decisions. And when the community does try to establish norms through model registries, responsible AI licenses, or voluntary safety frameworks, the norms are unenforceable because anyone can fork the model and ignore them.

The monitoring and sanctions conditions are the hardest. Who tracks what happens to a model after its release? The answer is no one, not comprehensively. There are partial monitoring efforts. Hugging Face has model card systems. Academic researchers audit specific models for specific failure modes. Security firms track known vulnerable fine-tunes. But there is no systematic monitoring of the global derivative ecosystem because the ecosystem is designed to be unmonitorable. That is the structural tension at the heart of open-weight AI. The property that makes it valuable for innovation, free access and permissionless modification, is the same property that makes it resistant to governance.

The US, the EU, and China are responding to this tension in completely divergent ways. The US is debating draft executive orders on open model risks while simultaneously benefiting from the innovation that open-weight models enable. The EU is building a regulatory framework that treats the most capable models as systemic risks requiring specific governance obligations, including disclosure and monitoring requirements for model deployers. China is releasing open-weight models like DeepSeek as part of its national AI strategy while simultaneously imposing domestic content controls on what the models can express. Each approach reflects a different answer to the same question Ostrom asked: can this shared resource be governed without destroying what makes it valuable?

I am not convinced any of the three has the answer. The US approach risks underregulating until a catastrophic incident forces reactive legislation. The EU approach risks overregulating in ways that push open-weight development to jurisdictions with lighter rules, which solves nothing globally. The China approach treats the commons as a state asset rather than a community-governed resource, which is precisely the centralized control model Ostrom's entire career was built on challenging. None of these approaches engages with the design principle question at the institutional level.

The platform governance literature in IS gives a more useful starting point. I wrote about how platforms govern through boundary resources: the rules, interfaces, and tools through which a platform owner coordinates participants while retaining control. Open-weight AI is a platform governance question in disguise. The releasing organization is the platform owner. The community of developers, researchers, and deployers are the complementors. The model weights are the boundary resource. The question is the same: who sets the rules, who monitors compliance, and who has the authority to sanctions violations. The difference is that the platform owner in open-weight AI releases the resource and then has very limited control over what happens next. Meta cannot inspect every fine-tune of Llama. Mistral cannot police every derivative model. The platform owner has distributed the resource but not the governance infrastructure needed to manage the commons it created.

Möhlmann, Gregory, and Henfridsson (2025) pushed this further by showing that platform governance is not only about owner control over complementors. Algorithms themselves shape the interactions and conflicts among stakeholders, and governance emerges from those interactions rather than being imposed from above. In the open-weight context, the interaction is between the model weights, the fine-tuning process, the deployment environment, and the end users who engage with the model outputs. The governance is not in the license. It is in the practices that develop around the model, and those practices are fragmented, inconsistent, and largely undocumented.

The IS field has the theoretical tools to address this problem. Ostrom gives us the design principles. Platform governance theory gives us the boundary resource framework and the stakeholder interaction model. Agency theory gives us the principal-agent lens for understanding who delegates authority to whom and with what accountability. The question is whether the field applies these tools to the open-weight problem before the governance failures become irreversible.

I think the binary debate between all-open and all-closed is where the policy conversation has been stuck, and it is the wrong framing. The question is not whether open-weight models should exist. They exist, and they are not going away. The question is what governance structures they need at the points in the ecosystem where governance can actually function: at the model release point, where the license terms and the model card specify what was done and what is permitted; at the deployment point, where the organization running the model accepts monitoring and reporting obligations; and at the incident response point, where a coordination mechanism must exist when a derivative model causes harm. These are governance design questions. They are institutional design questions. And the IS field has spent decades building the theory to answer them.

I am not sure the field is paying enough attention yet. Most IS research on AI treats it as a technology to be accepted or used rather than a common-pool resource to be governed. That framing misses the structural question. The open-weight AI debate is not about whether people will adopt these models. Hundreds of millions of downloads answer that question. The debate is about whether we can design the institutional arrangements that let the commons thrive without letting it collapse the way Hardin (1968) predicted, at a speed and scale Ostrom never anticipated. The field has the theory. The question is whether we show up.