The short version
AI policy is being captured at the level of epistemology. Not at the level of a single bill or a single agency. At the level of what gets counted as evidence, who is treated as an expert, and which questions are allowed to reach the table.
The entities with the most commercial interest in minimal AI regulation have become the primary shapers of AI governance worldwide. That is the paradox. It is named here deliberately. I am calling it the Spalding Paradox, because naming a thing is the first step toward being able to argue with it.
The fix is not new. It is older than any of the companies now sitting in the rooms where these rules get written. It is called the curb cut, and this post is about how we apply it to the hardest, most concentrated technology ever built.
Part one: the paradox
Here is the shape of it.
The companies with the largest economic incentive to prevent AI from being regulated are the same companies that have been invited to draft, redraft, advise on, comment on, and in many cases author the drafts of AI regulation. This is not a conspiracy. It is the mundane outcome of three overlapping enclosures that have all tightened at the same time.
The capability enclosure. Training a frontier model now requires compute budgets that only a handful of companies on earth can afford. Every other lab, every university, every nation state outside of maybe four of them, builds on top of someone else's foundation or not at all.
The trust enclosure. Governments, convinced by decades of market-first governance thinking, have concluded that the private entities closest to the capability are the most qualified to tell them what to worry about. Safety institutes and AI offices are staffed by people whose last job, and often next job, was at the same labs whose outputs they are now assessing.
The cultural enclosure. Public conversation about AI risk has been defined almost entirely by the frontier labs. What counts as a risk. What counts as reasonable. What counts as progress. The Overton window of AI policy narrows to the shape of a slide deck prepared by the companies that most benefit from a narrow window.
Each enclosure on its own is survivable. Together they compound. The people setting the rules are receiving their evidence, their framing, and their sense of what is possible from the entities that most benefit from those rules not being set.
This is why I call it epistemic capture rather than regulatory capture. Regulatory capture is when the regulated industry captures the regulator. Epistemic capture is when the regulated industry captures what the regulator can think.
The paradox is genuine, not rhetorical. If you build a regulatory system that takes expert evidence seriously, and the experts are concentrated in four labs, and the labs have commercial reasons to prefer minimal regulation, then taking expert evidence seriously is the mechanism by which the regulation arrives minimal. The good-faith act and the captured outcome are the same act.
Irony is when two things look alike and mean opposite things. Paradox is when the good-faith action produces the bad outcome structurally, not by accident. This is the second kind.
Part two: the curb cut that never got cut
The most widely cited accessibility win of the last fifty years is the curb cut. A single ramp at a pavement corner, built because wheelchair users demanded the city stop requiring them to perform a vertical gymnastic routine to cross the street.
The lesson everyone remembers is that the curb cut worked for wheelchair users. The lesson most people miss is what happened next. Parents with prams. Delivery workers with trolleys. Cyclists. People pulling wheeled suitcases. Kids on scooters. The curb cut built for one group turned out to be a structural improvement for everyone. The phrase for this is the curb-cut effect.
The curb cut worked because it was a permanent physical alteration to a shared space. Nobody had to ask permission to use it. Nobody had to identify as disabled to benefit from it. It cost almost nothing once the standard was written. And it became, over forty years, so thoroughly normal that most people under forty have never seen a pavement without one.
AI has not had its curb cut moment. Not yet.
The current landscape is the opposite of a curb cut. Access is permissioned. Use is metered. The populations with the most to gain, the neurodivergent, the disabled, the non-English speaking, the poor, the minimally verbal, are the populations least represented in the rooms where the standards are written. They are the last to receive the technology and the first to have it withheld on safety grounds that were defined without them.
"Safety", in the current AI conversation, is doing double duty. Some of what gets called safety is actual safety, and we should keep it. Some of what gets called safety is a commercial moat with a human rights PR strategy bolted to the front of it. Telling them apart is the work.
The curb-cut fix is to decide that AI is a shared public surface, like a pavement, and to design it from the corner outward. Not from the boardroom inward.
Part three: the K that becomes an 8
The economic term for the shape AI is accelerating is the K-shaped economy. The line of those who have capital, credentials, and connectivity slopes up. The line of those who do not slopes down. AI, without deliberate counter-design, makes the slopes steeper in both directions.
In the 8GI visual language, the K rotates forty-five degrees and becomes an 8. That is not a rebrand. It is a design target.
To make the K into an 8, the lower line has to curve back up. That does not happen by accident. It happens because someone builds the ramp. The curb cut.
The K-shaped economy is a diagnosis. The 8 is the intervention. The intervention is the curb-cut fix, applied at four points.
Part four: the four pillars
Pillar one. Open source as public infrastructure.
Frontier models that are entirely closed, entirely proprietary, and entirely under the commercial control of four US companies are, by construction, incompatible with a public-surface account of AI. You cannot curb-cut a pavement you do not own.
Open source models, with published weights, published training data where possible, and published evaluation methods, are not a charity position. They are the precondition of any governance worth the name. You cannot meaningfully regulate what you cannot inspect. You cannot meaningfully include populations whose needs you cannot verify are being met by weights you cannot read.
Every AI bill worth writing needs an open-source provision. Not as an exemption from safety rules, but as a route by which safety rules can actually be enforced.
Pillar two. Model access as a right.
The neurodivergent adult who cannot hold a job because no tool has ever been built to work with how their mind works does not need a subscription tier. They need the tool. The citizen who cannot afford to pay forty dollars a month for the most capable AI that their employed neighbours use daily does not need a discount. They need access as a default.
This is where subsidised compute, public-interest APIs, and sovereign national models enter. The analogy is not the telephone monopoly, where the state negotiates a lower price from a private operator. The analogy is the library. A place where the tool is available because a society decided it should be available, and because access is the thing, not the billing model.
Pillar three. Trust governance, not trust in governance.
Governance systems that rely on a small number of experts trusted absolutely are governance systems waiting to fail. The Spalding Paradox exists precisely because the experts most trusted are also the experts most incentivised against the outcomes the public actually needs.
Trust has to be manufactured structurally, not socially. That means disclosure regimes that apply equally to every lab above a compute threshold. It means civil society funding independent of the labs, so that the people capable of auditing are not the people employed by the thing being audited. It means safety evaluations that are adversarial, reproducible, and published.
And it means that when a national AI office is built, the conflict-of-interest rules are not discretionary. They are the first page of the founding statute.
Pillar four. Cultural access.
Regulation that emerges from one language, one cognitive style, one class of worker, and one national context cannot meaningfully regulate a technology that has already deployed into every language, every cognitive style, every class, and every national context.
The people most excluded from AI today are also the people least likely to be writing its rules. This is solvable. It requires budget lines for consultation that is not tokenistic. It requires accessible formats for consultation documents. It requires that the public does not have to be a policy professional to participate in policy.
I submitted to the Irish AI Bill in plain language. I did not submit in plain language as a stylistic choice. I submitted in plain language because policy written in a register only readable by people with law degrees is policy that chooses its participants before consultation begins.
Part five: why this is urgent now
There is a window, narrow but real, during which the rules of AI are being written. Most of them for the first time. The Irish AI Bill. The EU AI Act implementation phase. The US state-by-state patchwork. South Africa's national AI policy consultation. India's draft framework. A dozen others.
After this window, the rules harden. Not because they cannot be changed. Because the infrastructure that complies with them hardens, and the commercial interests that invested in compliance acquire a sunk-cost-shaped veto over future reform.
If the first cohort of AI rules is written under conditions of epistemic capture, the second cohort does not get written under better conditions. It gets written inside the footprint of the first.
The curb-cut fix is not a slogan. It is a scheduling claim. It says: the time to make AI a shared public surface is before the pavement is poured, not after.
The close
I am a full-stack engineer and the parent of a minimally verbal autistic child. I do not write about AI governance as a policy professional. I write about it because the outcomes of AI governance will shape whether my son gets to communicate at all, and whether the millions of people like him get to live in a country where the tools that could change their lives are available to them before they turn forty.
The Spalding Paradox is the structural trap. The curb-cut fix is the structural answer. The four pillars are where the fix gets built.
You do not wait to see how the game plays out. You cut the curb.
This post is companion to the interactive deck at /presentations/epistemic-capture, which carries the visualizations and audio chapters. The thesis will also be submitted to the South African AI Policy consultation and, in a revised form, to the Oireachtas follow-up round.
By James Spalding · Dublin · 8GI