The decisions being made right now about artificial intelligence are being made by a remarkably small number of people, in a remarkably small number of organizations, located in a remarkably small geographic area, drawn from a remarkably narrow demographic. The decisions they are making will determine the shape of labor markets, information environments, social organization, and the distribution of power for whatever decades remain in front of us. The people who will bear the costs of these decisions are not the people making them. The architecture by which those costs would be negotiated by the affected populations does not exist.
Within the last four months, two of those operators have publicly reversed the most structurally specific statements they had made about the labor consequences of what they are building. The reversal coincides with their firms’ preparations for IPOs estimated at one trillion dollars each. The data that would justify the magnitude of the reversal has not materially changed. The operators have changed.
This essay is about the structural condition that produces both the original statement and the reversal, the question of what the historical record suggests about transitions of this kind, and the limits of what any analysis from the position I am writing from can claim.
Care is required in how this is laid out. The diagnostic apparatus I have been developing, a framework for identifying coherence and capture across organizational and political scales, maps to this condition with such ease that I do not entirely trust the mapping. Explanatory monism, the failure mode of an analysis that fits every observed condition and therefore explains nothing in particular, is real, and the discipline of this work requires me to name it. The structural diagnosis available here is available because the structure is genuine. It is also available because the diagnostic was built inductively from the same kind of pattern. I will mark the limits where I see them.
The structural facts are not contested.
AI development at frontier scale is concentrated in a handful of organizations. Decision-making about pace, capability, deployment, and risk mitigation is held by a small number of capital holders and executives within those organizations. The financial upside of the technology, to the extent it produces upside, accrues to those holders in the first instance, with downstream distribution determined by political and economic infrastructure that, in the current moment, has been substantially weakened. The financial and social downside, to the extent the technology produces downside, falls on populations that have no representation in the decisions and no architecture through which to participate in them.
The diagnostic principle applies cleanly. The stated purpose of frontier AI development is the benefit of humanity. The actual operation is unprecedented capital concentration, the redirection of cognitive infrastructure toward systems whose deployment patterns are determined by commercial pressure, and the construction of a labor-displacement vector whose magnitude the decision-makers have publicly disputed with themselves within the span of a single year. Who benefits from the distance between the stated purpose and the actual operation is not a difficult question.
What is harder is a distinction the analysis requires me to preserve. The capital holders, the executives, the senior researchers, the working engineers, the contract data labelers, the support staff, the deployment teams, the safety teams, the policy teams. These populations are not identical. They do not share interests. Operators must be distinguished from captured at every scale of analysis, and must be distinguished here. Most of the people working in AI are not the operators of the disconnect. They are operating inside an architecture that has captured them in some of the same ways the architecture has captured the populations they are displacing. Some understand the structural position they occupy. Some do not. Some are accelerationist about it. Some are working from inside to alter the trajectory. Some are doing the technical work that makes the systems possible and are aware, in varying degrees, that the systems will be deployed in ways they cannot fully control. The diagnosis is sharper, not blunter, when these distinctions are preserved.
This matters because the broad framing of oligarchs creating AI obscures the structural reality of who decides and who builds. The operators are a much smaller group than the builders. The builders include populations who, under different architectural conditions, could be the counterweight rather than the substrate.
The historical pattern, where this kind of transition has occurred before, is consistent enough to be diagnostic without being universal.
During every major technological transition in the modern industrial era, the operators captured the architecture first. The counterweight infrastructure was built later, slowly, at enormous cost, by movements that had to assemble themselves under conditions the operators had already constructed.
The Gilded Age robber barons consolidated railroad networks, oil refining, steel production, and finance into trusts that operated outside any effective regulatory constraint for roughly four decades. The counterweight, when it eventually emerged, came from the Progressive movement, from the labor movement, from the muckrakers who documented the architecture for the public, from the antitrust legislation that took the better part of a generation to assemble, and from the New Deal architecture that finally provided structural constraints capable of limiting concentrated capital. The constraint did not arrive in time to prevent the harm. It arrived in time to limit the next iteration of the harm. The first iteration ran its course at the cost of generations.
The industrial revolution displaced rural labor into urban factories under conditions whose architecture was set entirely by capital. The counterweight, the labor movement and the eventual administrative state, was built over a century, against violent resistance, at the cost of large amounts of human destruction. The eventual architecture was real and produced genuine constraint. The constraint did not undo the displacement. It produced different terms for the next set of displacements.
I am selecting these cases because they are the cases most structurally similar to the current transition. They are not the only cases of large-scale architectural shift. Spain’s Pacto del Olvido, Estonia’s institutional reconstruction after Soviet collapse, post-Khmer Rouge Cambodia, post-genocide Rwanda, each is a case of architectural transition under different conditions, and each would complicate or refine the pattern I am about to extract. I am not engaging with them here because the question of this essay is the structural condition of operators capturing architecture in advance of counterweight, and the Gilded Age and industrial cases are the canonical instances. The selection is a limit on the analysis that I am naming rather than concealing.
The pattern, where it holds, is not optimistic. Operators build the architecture in the absence of constraint. Constraint, when it eventually emerges, arrives late, partial, and through enormous external pressure. The pattern is not an inevitability. It is the default trajectory in the absence of architectural intervention.
The question for the current transition is whether the default trajectory will produce the same eventual counterweight, or whether the current transition has structural features that make the historical pattern an unreliable guide.
The evidence that this transition is structurally different has been supplied, in considerable detail, by the operators themselves. The evidence has also, within the last several weeks, been substantially walked back by the same operators. Both moments matter. Neither neutralizes the other.
In January 2026, Dario Amodei, chief executive of Anthropic, published an essay titled The Adolescence of Technology, the fourth section of which addressed labor market disruption directly. The essay was, at the time of its publication, the most structurally specific public statement any frontier AI executive had made about the labor consequences of the technology they were building. It identified four structural features distinguishing this transition from prior technological transitions. The pace of capability gains is faster than the rate at which workers and labor markets can adapt. The technology’s reach across cognitive abilities is broad rather than narrow, eliminating the path by which past transitions allowed workers to migrate from one cognitive domain to another. AI development proceeds from lower to higher cognitive abilities in a way that targets populations by intrinsic capacity rather than by retrainable skill, producing the risk of a permanently unemployed underclass. The technology adapts to its own gaps faster than past technologies, which closes the buffer that historically allowed humans to fill the remaining work that the new machine could not yet do.
The prediction Amodei attached to this analysis was that AI could displace roughly half of all entry-level white-collar jobs over a one-to-five-year horizon. He was explicit that this was not a refusal of the law of comparative advantage in principle, but a recognition that comparative advantage breaks down when productivity differences become large enough that transaction costs make the trade not worth pursuing. He acknowledged that historical analogies provided limited guidance. He described the economic concentration of power already in formation: the largest individual fortune in the world today already exceeded, as a fraction of GDP, the Rockefeller fortune at the peak of the Gilded Age, before most of the economic impact of the technology had been delivered.
The defenses Amodei proposed in the same essay were real. Better real-time data. Steering enterprises toward innovation over cost-cutting. Companies treating their displaced workers with care. Private philanthropy. Government intervention through progressive taxation. He was explicit that the voluntary actions of any single firm could not solve the externality, and that the macroeconomic problem would require government action. He was also explicit that he was making this argument under conditions in which the political infrastructure that would normally translate such an analysis into policy had been substantially degraded, and that the firms that would have to participate in the construction of the defenses operated under competitive pressure that made voluntary restraint structurally unstable.
In May 2026, four months later, Amodei reframed automation, in public statements, not as a destroyer of jobs but as a multiplier of output. If you automate 90% of the job, then everyone does the 10% of the job, and the 10% kind of expands to be 100% of what people do and kind of 10-times their productivity. The frame is Jevons paradox: as automation reduces the cost of cognitive labor, demand for it expands, and the workforce is absorbed by the expansion rather than displaced by the substitution.
Sam Altman of OpenAI, in the same week, told an interviewer he had been pretty wrong about AI’s economic impact. The displacement he had feared, he said, had not materialized. A personal experiment, in which he tried delegating email to AI and found himself preferring to respond manually, had updated him toward a view in which the jobs picture would be different from what he had feared.
David Solomon of Goldman Sachs, in a New York Times op-ed published the same week, did not need to reverse position because he had never held the apocalyptic view. He drew a straight line from electrification in the 1900s to the digital revolution of the 1990s to the present, citing growth in civilian employment of 145% since 1962 and the 200,000 jobs Goldman Sachs research attributed to data center construction since 2022. Other technology CEOs and economists offered variations of the same long-run argument.
Both OpenAI and Anthropic were reported, in the same window, to be preparing IPOs at an estimated valuation of one trillion dollars each.
This is the moment where the framework I have been articulating most wants to declare itself vindicated, and where the discipline of the work requires me to be most careful in both directions.
The honest reading requires holding two things together without letting either neutralize the other.
The alternative explanations are real. The displacement curve Amodei described in January was a prediction. Predictions can be wrong. The Yale Budget Lab has found no significant changes in occupational mix or unemployment duration in high-AI-exposure jobs since ChatGPT launched in late 2022. The original predictions had a one-to-five-year horizon, of which roughly three years have elapsed. Absence of effect at year three does not refute predicted effect at year five, but it does provide a real evidentiary basis for downward revision. Jevons paradox is a legitimate economic dynamic. The argument that automation expands rather than contracts demand for cognitive labor is one that many serious economists have made about prior technological transitions, with mixed historical support: some transitions produced eventual labor absorption, others produced lasting displacement that took generations to resolve and produced enormous medium-run human cost. These considerations have weight. I will not pretend they do not.
The structural facts are also real. Nothing in the data materially changed between January and May. The Yale Budget Lab data was available in January. The Jevons paradox argument was available in January. Solomon’s long-run argument was available in January. The displacement curve Amodei described in January was constructed in full awareness of these counterarguments. He explicitly addressed comparative advantage and explicitly acknowledged the historical pattern of labor absorption, and concluded, in January, that this transition was different. What changed in May was not the data. What changed in May was the constituency the operators were addressing. In January, the audience was the policy community and the political infrastructure that needed to be reassured that the firms were taking the labor risk seriously enough to justify continued political accommodation. In May, the audience was the capital infrastructure that needed to be reassured that the workforce would absorb the technology rather than reject it, and that the firms’ valuations did not depend on a labor disruption that would invite political constraint. Same operators. Different audience. Different statement.
The position from which Solomon offers his vindication is not neutral either. He is the chief executive of one of the firms with substantial financial exposure to AI infrastructure expansion. The 200,000 jobs in data center construction he cites are jobs his firm has financed. The long-run argument he makes, in support of the framing that AI will not produce significant labor disruption, is the argument that serves the financial interests of the firm he leads. This does not make the argument wrong. It does make the position from which it is offered part of the data.
The honest synthesis is that both dynamics are present. The operators may have genuinely updated on early evidence. They are also managing constituencies whose interests in the framing diverge, and the timing of the reversal aligns with the constituency shift in a way the diagnostic principle requires us to mark. The structural pattern, the disproportion between the magnitude of the reversal and the actual change in evidence, coupled with the alignment between the reversal and the financial event, carries enough weight that treating the reversal as a clean epistemic update would itself be a failure of analysis. Treating the reversal as definitive evidence of capture, without acknowledging the genuine evidentiary basis for revision, would be a different failure of analysis. Both moves are available. Both are wrong. The work is to hold the synthesis without collapsing into either.
What can be claimed about this is narrow but specific.
The structural condition is a capture in formation. The architectural decisions about what AI will do, at what pace, with what mitigations, to which populations, are being made now, by a small number of actors, inside an environment where the political infrastructure that would historically have negotiated the terms of such a transition has been substantially degraded. Labor power is at a historic low. The regulatory state in the United States is operating under conditions of active dismantlement. The media architecture has been substantially captured. The traditional civic intermediaries that brokered past transitions, the unions, the professional associations, the watchdog institutions, are operating with degraded capacity. The architecture that, in past transitions, eventually constrained the operators is largely not available.
I cannot predict whether AI will displace half of entry-level white-collar jobs or expand the work people do or land somewhere between. The analysis can identify the structural condition under which that question will be settled. The condition is: a small number of actors making consequential decisions, inside an environment that has lost the infrastructure to constrain them, with no procedural mechanism for the affected populations to participate. That structural condition, generalized across past transitions of this character, has produced outcomes that meet the definition of architectural capture. The operators set the terms. The substrate bears the costs. The eventual reckoning arrives late and at enormous cost.
The reversal documented in the previous section is itself an instance of the structural condition operating in real time. The operators can revise their public position about the magnitude of the disruption they are constructing without consulting the populations who will bear the costs of being wrong about it. No mechanism exists by which those populations can participate in the framing decisions that determine whether they will be treated as a constituency requiring protection or as a workforce that will absorb the technology smoothly. The framing is itself a unilateral operator decision. The absence of counterweight is what makes the framing unilateral. The reversal is not, by itself, evidence that the original prediction was wrong. It is, by itself, evidence that the framing of the labor question is being conducted in the absence of any architecture that would constrain the framing.
The architecture being built right now is being built without the counterweight that would, in a coherent system, constrain it.
The constraint will not come from inside the firms doing the building. The firms cannot constrain themselves under competitive pressure, a point the leading firms have stated publicly and continued to state even while reversing the public framing of the disruption their products will produce. The constraint will not come from the workers inside the firms, most of whom lack the structural position to override commercial decision-making. The constraint will not come from the affected populations, who do not have the institutional infrastructure to assemble counterweight in the available window. The constraint will have to come from architectural intervention that does not currently exist and is not currently being built.
What can be said about the current moment is narrow but specific. The capture is not complete. The architecture is in formation. The infrastructure of counterweight could be built in the available window if the architectural decision were made to build it. The decision is not being made. The reasons it is not being made are the reasons capture proceeds at every other scale of analysis. The operators benefit from its absence. The would-be architects of constraint lack the institutional position, or the will, to overcome that benefit. The political infrastructure that would historically have brokered the construction has been substantially degraded. The populations who would bear the costs are not organized into a form that can participate in the decisions being made about their lives.
Care is required in the closing claim, because it is the place where consolation would be a form of dishonesty, and the place where the position from which the analysis is delivered must be named again.
There is no prediction here that the counterweight will be built. The prediction is more limited. In the absence of architectural intervention, capture proceeds to the conclusion the architecture incentivizes. The current trajectory is the trajectory the architecture is producing. Whether the trajectory will be interrupted is not a question this analysis can answer. It is a question of whether the people who could build the counterweight will do so, under conditions that make the building extremely difficult, in time to matter.
The counterweight has to be built by people who are not in this position. What documentation can do is be a record that survives the moment. It can name the structural condition with enough precision that the architects of constraint, when they arrive, have a description of what was built and what was being said about it at the time it was built. It can hold the position of the operators visible at the moment of the constituency shift. It can refuse to let the reversal pass into the historical record as a clean epistemic update when the structural evidence requires it to be marked as the position-shifting that capture in formation produces. It can also refuse to let the reversal stand as proof of capture when the genuine evidentiary basis for revision is what it is. The documentation holds the synthesis.
The disconnect is being constructed now, by people who do not bear its costs, inside an environment that has lost the architecture to constrain them. The construction is consistent with historical patterns of capture during technological transitions, and was, on the original testimony of the operators themselves, more aggressive in structural respects than past transitions have been. That testimony has been substantially walked back. The walking back has not been accompanied by material new evidence. The conditions for the historical counterweight to assemble are weaker than they have been in a long time. The structural diagnostic is identified. The actor who will respond to it is not.
What I can do is document what is being built, what is being absent, what is being said about both, and what the structural condition predicts in the absence of intervention. The documentation is what I can provide. The intervention is for others.
Jason I. Oh is the author of The Architecture of Coherence: A Systems Theory of Organizational Justice and the originator of the topological realism framework.