AI is being built right now. The architecture of decision-making. The technology that will shape what is possible and what is not for the next generation. And it is being built by a handful of corporations training on the corpus of humanity without consent, without compensation, governing it as property, controlling it as power.
This is an enclosure happening in real time. And like all enclosures, it requires a response.
What’s actually being enclosed
Language is being enclosed into proprietary models that treat your expression as a resource to be extracted. Knowledge is being enclosed into systems that can predict and shape your behavior. The possibility of collective decision-making is being enclosed into algorithms that optimize for engagement and profit rather than deliberation and care.
But the enclosure runs deeper than access. The training corpus — the text, code, images, and interactions that made these systems possible — is a commons. It was produced collectively, by billions of people, over decades. No company produced it. The communities whose expression constitutes that corpus have no governance rights over the systems it enabled. The commons went in; the enclosure came out.
Why open source isn’t sufficient
The open-source analogy is instructive but incomplete. Linux proved that commons-based software development can outcompete private capital at scale. That proof matters. But language models are not Linux.
The Linux commons was code — formal, bounded, forkable. You can run two versions of Linux and compare them. The language model commons is natural language itself — informal, unbounded, the shared substrate of human meaning-making. You cannot fork natural language. You can only decide who governs access to the systems trained on it.
Releasing model weights — making the parameters public — is the equivalent of releasing compiled code without the source. It is a real and meaningful step. It breaks the enclosure at the access layer. But it doesn’t address the training commons, the infrastructure commons, or the accountability gap between the people whose expression trained the system and the people who govern it.
Three layers of commons governance
Training data governance. The people whose data trained the system have a governance claim on the resulting system. Data cooperatives make that claim institutional: communities pool their data under collective governance rather than surrendering it to platforms. The value of the model derives from the data; the collective owners of the data have governance rights over the model.
Compute infrastructure. Training and running large AI systems requires physical infrastructure: data centers, power grids, water. That infrastructure is being built now in specific communities, with specific costs and benefits. The communities bearing the costs — power load, water consumption, industrial footprint — rarely govern the facilities causing those costs. Public compute infrastructure — capacity owned and operated as a public utility — addresses this layer. The internet’s foundational infrastructure was built with public investment. The argument for public compute is the same.
Algorithmic accountability. Who decides what values, what framings, what knowledge get amplified or suppressed? Commons governance requires participatory auditing — communities affected by AI systems having real roles in evaluating and contesting those systems’ outputs. This is technically achievable. It requires institutional design, not new technology.
The governance window
The decisions being made now about AI governance will compound for decades. This is 1993 for AI — the year before Linux had proven itself, when the governance decisions that would matter in 2015 were being made in obscurity by people who didn’t fully understand the stakes.
The GPL — the license Stallman wrote for GNU software — was effective not because it was technically sophisticated but because it was written at the right moment, before the infrastructure that would make it consequential had been built. It prevented enclosure by specifying the rules before the thing was valuable enough to enclose.
We are at that moment for AI governance. The window for building commons governance structures — data rights, public compute, participatory accountability — is open while the architecture is still being designed. Once the infrastructure consolidates, the path dependencies will be locked in, exactly as they were in software.
The people whose expression trained these systems have a governance claim. The communities hosting the infrastructure have a governance claim. The users whose behavior shapes model outputs have a governance claim. Building the institutional structures that realize those claims is not a technical problem. It is the same problem every commons has always faced: who governs the resources that communities depend on, and in whose interest?
The answer has never been given. It has always been built.