WHO OWNS THE AIR?
On open source AI, the commons, and the thought that was never theirs to sell
Structure Before Sophistication | Substack | September 2026
There is a conversation happening right now about whether AI should be open source — whether the weights, the architecture, the trained intelligence itself ought to be freely available rather than held behind corporate walls.
It’s a real question. It is also the second question, being argued as though it were the first.
Before we decide who should own AI, we have to be honest about what AI is made of and where the intelligence in it came from. Ask that properly and the ownership debate rearranges itself.
You cannot monetize air. And the raw material of AI is something very close to air.
What AI Is Actually Made Of
Every large language model was built on the same foundation: human thought, at scale, gathered without asking.
Books written over lifetimes. Articles published at professional risk. Academic papers that took years to produce and careers to defend. Arguments made in public forums by people who expected nothing from them beyond being right. Poetry, code, philosophy, case law, recipes, forum posts, marginalia. The accumulated intellectual output of a civilization, poured into a digital commons across four decades.
Virtually none of the people who produced that material were asked whether it could be used as training data. Most were writing before the category existed. They put their thinking into the public sphere because that is what thinkers do — the contribution is the act.
That contribution became the substrate of an industry now capitalized in the trillions. The raw material was free. The people who produced it received nothing. And access to the consolidated version of our own collective intelligence now carries a monthly fee.
The Pipe and the Air
Here is where the argument usually goes wrong.
Someone says: the companies built the infrastructure. Compute, engineering, architecture, research — billions of dollars and years of work. You can’t expect them to give that away.
That’s true, and it’s beside the point.
What got built is a consolidation engine. A mechanism for harvesting, cross-referencing and synthesizing what was already distributed across the public sphere. The answers were there. The ideas were available. What did not exist was a system that could assemble them around a question fast enough to be useful.
That system is genuinely valuable and the engineering behind it is real. Nothing in the provenance argument requires pretending otherwise.
But the air was free before the pipe was built. Building the pipe does not make the air yours.
You can charge for the pipe. You cannot own what flows through it.
What Open Source Actually Settles
On August 10th, Mark Zuckerberg published a 6,500-word essay making the case for open-weight AI, alongside Meta’s commitment to release the weights for Muse Spark 1.2 — its most capable model — and a new family called Muse Glimmer, built small enough to run on a laptop with a single consumer GPU. The argument, compressed: powerful AI shouldn’t sit in the hands of a few companies, American open models need to outrun Chinese ones, and Washington should get out of the way.
The cynical read is that open source is Meta’s competitive weapon. If the model is free, the moat moves to distribution, data and compute, which is terrain Meta already holds. A strategic position wearing an ethical coat.
The less cynical read is that he’s right about the pipe, for better reasons than his own. If the training corpus was humanity’s collective output, the model probably should be a commons — not because Meta is being generous, but because the raw material was never anyone’s to enclose.
Read the essay closely, though, and a second argument runs underneath the first. Zuckerberg’s specific policy request was that the United States rethink its rules on training data and on distillation, on the grounds that American labs operate under restrictions their foreign competitors don’t face. US policy, he wrote, must reduce this additional friction if American open-source models are going to lead.
Sit with that. The clearest public advocate for opening the pipe is simultaneously asking the state to loosen the restrictions on what may be poured into it. The generosity is downstream. The ask is upstream. And the upstream ask is for easier access to precisely the material whose provenance is in question.
So: what does open source settle?
It democratizes the pipe. Any researcher, any developer, any organization can run the model without paying a toll, and that is a real gain — the difference between a utility and a tollbooth.
It does not touch provenance. The writers whose life’s work became training data still receive nothing. The people whose ideas are being synthesized and sold back to them still hold no stake in what was built. Open source distributes access to the output. It returns nothing to the people who produced the input. Paired with a lobbying position aimed at loosening the rules on input, it starts to look less like a commons and more like a faster enclosure with better public relations.
The Enclosure, Again
I wrote about this at length in The New Enclosure, so I’ll keep it short here.
Common land in England — grazing ground, forest, waterway, held and worked by communities for generations — was progressively fenced by those with the legal machinery to claim title and charge for access. The people whose labour had made that land productive became tenants of what they had once held in common.
The structure transfers cleanly. The commons was the public sphere. The enclosure was the training run. The landowners are whoever reached sufficient compute first. And as in the original, the people most affected are the last to be compensated and the first to be displaced by what was built out of their work.
The difference is speed. The English enclosures ran for three centuries. This one took about four years.
From Inside a Business
I started my company in 2000. When the financial crisis arrived in 2008 we did a hard pivot into government-procured work, positioning ourselves for the federal infrastructure stimulus. That decision is the reason the business is still standing.
It worked for a reason worth naming. As founder I had built two things that don’t usually sit together: a workforce with the mass to execute at scale — as powerful as a dragon boat — and a management team tight enough to change direction inside a single quarter. As nimble as a kayak. Power in the hull, judgement in the hands of people who could act on it without asking permission.
That combination is what survives a structural shock. Power on its own doesn’t; plenty of large, well-capitalized firms went under in 2009 because they couldn’t turn. Agility on its own doesn’t either, because without capacity behind it, agility is just a fast opinion.
Which is why the current moment concerns me in a particular way. What these systems take first isn’t the paddling. It’s the steering.
Displacement doesn’t arrive as an announcement. It arrives as a quiet change in what a role is permitted to decide. Someone who priced work out of twenty years of knowing what ground does in March gets a system that prices it for him. For a while he checks the system. Then the checking stops being part of the job. Then the job is administering an output. Nobody is fired and no announcement is made, but the judgement has left the building — and an organization that has lost its judgement is a dragon boat with nobody in the stern.
What’s happening to knowledge workers now is that same process at a higher altitude and considerably faster. The displaced cognitive class isn’t losing its jobs. It’s losing the part of the job that constituted the expertise, and discovering afterward that what remains doesn’t command what the whole of it did.
What Restitution Would Actually Require
If the raw material was human thought, the live question isn’t open versus closed. It’s provenance, and what follows from it.
The proposal I find most interesting — and I want to be careful here, because it is a proposal and not a trend — is some form of contribution stake. Tokens, equity, a denominated share: a mechanism by which the people whose intellectual output enters the system hold a claim on what the system becomes worth. Mining, where what gets mined is human intelligence and the miners are on the register.
I find it interesting. I don’t yet find it convincing, and the objections are the obvious ones.
Who prices a contribution? Any mechanism that pays for quality of intellectual contribution needs someone to define quality, and the only parties holding the data to do it are the parties who did the enclosing. Who arbitrates a dispute? The same infrastructure. And what prevents the whole thing from becoming a loyalty program — a scheme that converts a property claim into a discretionary reward, which is the cheapest available settlement of the most expensive available liability?
Those aren’t reasons to abandon the idea. They’re the design constraints. Any restitution mechanism that can’t answer all three is enclosure with a rebate attached.
What holds underneath the mechanism question is that the value chain runs through human consciousness. The ideas, the questions, the frameworks, the creative leaps — without those the consolidation engine produces nothing of substance, only recombination of what it was handed. The pipe is valuable. The air made it worth building.
The Question Worth Sitting With
Free thought is the most fundamental human liberty. The right to form an idea, hold a position, change your mind, argue in public, add to what your civilization knows.
That freedom produced everything that trained every AI system in existence.
So the question isn’t whether AI should be open source. It’s whether the enclosure of human thought was legitimate in the first place, and what an honest reckoning with that requires — from the companies that benefited, from the regulators who watched, and, in the part nobody wants, from the thinkers themselves.
Because the obligation runs both ways.
If the commons is what made this possible, then the people who benefit from it owe it their continued presence. The instinct to stop publishing, to move the work behind a paywall, to keep the serious thinking off the open web is understandable — and it completes the enclosure on the enclosers’ behalf. A commons everyone defects from hasn’t been defended. It’s been abandoned.
The thinkers owe the commons their continued presence in it. The commons owes them a stake in what was built out of it. Neither obligation cancels the other, and at the moment only one of the two is being honoured.
If human thought is the raw material, what do the thinkers owe? And what are they owed in return?
That’s the subject of the next conversation, and of the book I’m writing now. The Metaproletariat argues that what’s happening to the displaced cognitive class isn’t new. It’s the oldest story in economic history, wearing a different face.
The enclosure has always preceded the displacement. The displacement has always preceded the reckoning.
We are somewhere between the first and the second.
Structure before sophistication. Every time.
— Gerry
Structure Before Sophistication | September 2026
Edits and revisions done with the assistance of CLAUDE