STPA Analysis of Frontier AI as a Sociotechnical System¶
What Will AI Do to Us? A Systems-Engineering Reading¶
Abstract¶
The diffusion of frontier artificial intelligence is reshaping the global economy and political order, but most public analysis of this fact takes the form of strategic essays — narratives that identify forces and rank scenarios without pointing at specific levers. This chapter applies System-Theoretic Process Analysis (STPA), with justification-rung tagging from the social-systems extension developed in Part I, to the global political economy of frontier AI. We identify fourteen distinct controllers, twenty-six control actions, sixty-five unsafe control actions across the four standard categories, and nine loss scenarios connecting specific structural mismatches to specific harms. We then derive twenty-two concrete remedies organised into a four-phase sequence with explicit decision points. The result is a structural reformulation of the questions about AI's effects on jobs, power, and geopolitics — turning forecasting questions into action-shaped questions of which feedback channels are missing and which institutions are needed to install them. The full analytical artefacts are filed under projects/problems/ai-impact-analysis/; this chapter distils them into prose for a non-specialist reader.
1. The questions, and why the usual answers do not satisfy¶
Three questions about artificial intelligence circulate in every serious conversation about the next decade.
The first is about jobs. If software writes itself, what becomes of the people who wrote it? If diagnoses, contracts, lesson plans, research summaries, and software designs are produced in seconds by machines that cost a fraction of a human salary to operate, how many human roles survive? Will the result be a new wave of unemployment, or a new kind of leisure?
The second is about power. The systems that now demonstrate near- human reasoning at language tasks were built by perhaps a dozen laboratories, financed by perhaps a hundred investors, and trained on perhaps a few hundred specialised computers in a few jurisdictions. If the technology becomes the central productive force of the next era, what stops the firms that own the substrate from owning, by extension, much of what the substrate produces?
The third is about geopolitics. The same technology is being built in two large economic blocs that mistrust one another. The American bloc's frontier laboratories sit at the top of the global capability ranking; the Chinese bloc has matched them at lower rungs and is catching up at the top. Export controls, sanctions, and chip-supply dependencies have already produced the first skirmishes of an infrastructure war. Where does this end?
Each question has been answered many times. The answers are sophisticated, but they share a common shape and a common limitation. The shape is the strategic essay: a narrative that identifies the main forces, ranks them by importance, and concludes with a list of probable scenarios or a cautious warning. The limitation is that the essay's conclusions, however confidently stated, are free-floating. They are not connected to specific mechanisms in the world that one could intervene on. The essay tells the reader what is likely to happen; it rarely tells the reader where the system can be moved.
This chapter does something different. It treats the AI economy and its political surroundings as a system in the engineer's sense: a collection of parts connected by flows and feedback loops, each of which can be examined and, if it is part of a failure pathway, changed. The discipline that does this is called systems engineering, and a particular variant of it called system-theoretic process analysis (STPA) was originally developed for safety-critical machines like aircraft and medical devices. We use it here for a different class of artefact: not a machine, but the global political economy of frontier AI.
The reader gains, by the end, a structural picture of the system, five specific identified failure mechanisms inside it, nine specific scenarios in which those mechanisms produce identifiable harms, and twenty-two concrete interventions arranged into a four-phase implementation plan. This is more than an essay can deliver, and the reason it is more is the method. So we begin with the method.
2. Two ways of thinking about the system¶
A strategic essay about AI says things that sound like this. Power will concentrate, but open-weight models from China will keep prices honest. Workers will be displaced, but new categories of work will emerge as they always have. The geopolitical risk is significant but probably manageable. The most important uncertainty is whether redistribution mechanisms catch up to the productivity shock. These sentences are useful for orientation. They are also, on closer examination, the kind of observations that one could make about almost any technology in almost any decade. They do not tell you which lever to pull. They do not even tell you which levers exist.
A systems-engineering analysis, by contrast, begins by asking who is actually in the system, what each of them controls, and what information each of them receives. It records this not as a narrative but as a small set of explicit lists: the controllers, the things they control, the control actions they issue, the feedback channels they receive, and — crucially — the internal model each controller carries of the thing it is controlling. It then asks, for each control action, what it would mean for that control action to be issued unsafely (provided when it should not be, withheld when it should not be, given too late, given for too long). Each unsafe possibility is recorded. Then it asks, for each unsafe possibility, why the unsafe action could occur — the controller's model could be wrong, the feedback could be missing or filtered, the action could be misexecuted, the controlled process could have changed underneath. Each cause suggests a remedy: fix the model, fix the feedback, fix the actuator, fix the structure.
This sounds like ordinary strategic thinking with extra paperwork, and at first glance the difference is just that there are tables where the essayist has paragraphs. The deeper difference shows up at three points. First, the act of writing down every control action and every feedback channel forces the analyst to confront parts of the system that the essayist could quietly skip past; the omissions become visible. Second, recording each unsafe control action against the four standard categories produces a checklist of things-that-could-go-wrong that is exhaustive in a way that strategic intuition is not. The analyst is no longer free to remember only the failure modes that fit the narrative. Third, the loss-scenario step demands that the analyst not just assert that a failure could occur but trace the causal chain through the structure. This is what produces remedies. A narrative observation that "regulators are slow" is not actionable; a loss scenario showing that labour ministries operate on a years cycle while capability deployment happens on a months cycle, so the corrective signal arrives after the cohort it was meant to help is already displaced points directly at automatic stabilisers indexed to a fast-moving benchmark.
There is a further refinement, particular to systems engineering applied to social and political systems: every flow in the structure can be tagged with the justificatory rung on which it operates. The rungs are a seven-step ladder of how a claim or a control action gets its force. Rung 0 is naked coercion: do it or else. Rung 1 is authority: do it because I say so, or because the tradition says so, or because our group says so. Rung 2 is formal consistency: this follows from rules we both accept. Rung 3 is empirical evidence: this works against rivals when tested. Rung 4 is replicated cumulative evidence. Rung 5 is meta-rational integration of multiple methods under uncertainty. Rung 6 is deliberative legitimacy — the standard by which questions about what should be done are settled in open discourse among those affected.
Each of these rungs can be claimed. Each can be operated. The two are not always the same, and where they differ, the system typically produces specific kinds of damage. A laboratory that claims to operate at rung 6 (its safety framework is "for the benefit of humanity") but actually operates at rung 1 (the framework binds only the laboratory itself, by its own choice, amendable by its own governance) is in a claim/operation gap. A funding flow that goes downward at rung 3 (financial-return diligence) but receives capability evidence upward at rung 1 (the lab's own benchmarks, in the lab's own framing) is in an asymmetric loop — the loop appears closed because there are arrows in both directions, but the upward arrow is filtered by the downward arrow's framing, and corrective information cannot get through.
Once you start labelling every arrow with its rung, three patterns keep appearing. The first is the asymmetric loop just described. The second is the claim/operation gap. The third, which I will explain when we encounter an instance, is cross-loop rung imposition — when an institution tries to settle a question that belongs at one rung using the authority of a different rung. Each of these patterns has a generic remedy that the engineering literature catalogues. The remedies are structural — they change the architecture, not the personalities — and that is what makes them robust.
We now have everything we need to look at the AI system itself. We will list its parts, draw its loops, mark the mismatches that the rung-tagging method surfaces, walk through the failures these mismatches produce, and end with the remedies they imply. The reader who finds the technical vocabulary unfamiliar can lean on the prose: we will introduce each new term in plain English at the moment it enters the chapter.
3. The system¶
To draw the AI system's control structure honestly we have to be willing to write down a longer list of actors than a strategic essay would normally bother with. The longer list matters because the failures we are looking for happen in the connections between actors, and a connection is invisible if either of its endpoints is missing from the picture.
At the apex sits something we will call the global polity: diffuse, slow, but the source of legitimacy on which most controllers ultimately depend. Below it, three large state groupings — the United States, the People's Republic of China, and the European Union together with other states — each with several distinct organs. The United States in this analysis is not one actor but six: the Congress that passes legislation, the Bureau of Industry and Security that issues export controls, the Treasury that runs sanctions and tax, the Federal Trade Commission that runs antitrust, the AI Safety Institute that runs technical evaluation, and the national security apparatus that classifies and designates. The People's Republic of China is three: the party centre that sets strategic direction, the Ministry of Industry and Information Technology that runs industrial policy, and the Cyberspace Administration that regulates content and approves models. The European Union is two: the Commission with its AI Office, and the member states. Other states — the United Kingdom, Japan, South Korea, India, the Netherlands, Taiwan, the Gulf monarchies — appear as a coalition layer whose compliance or defection determines whether the bigger states' levers actually move anything.
Beneath the state layer sit the capital markets in their various forms — public equities, private capital, sovereign wealth, debt. These are the controllers that decide which laboratories can sustain frontier training runs and which compute providers can sustain capital-expenditure cycles. Beneath them sit the frontier laboratories themselves, perhaps a dozen worldwide, each with its internal mission narrative, its internal capability assessments, and its internal safety framework. Beneath the laboratories sit the compute infrastructure providers — the chip designers, the foundries, the hyperscale cloud operators, the energy and grid companies. To one side of this hierarchy, neither below nor above the labs but functionally parallel to them, sits the open-weights ecosystem: the laboratories, mostly Chinese but including some Western publishers, that release model weights free for download. On the other side sit the application-layer companies that turn model capability into deployed software for paying customers. Below those, the information platforms that distribute everything. Below them the labour ministries and education systems that interface with the people whom all of this affects, and the information regulators that look after the integrity of the public information environment. At the bottom — though "bottom" is wrong because they are the controlled process this whole apparatus ultimately produces effects on — are the workers, consumers, and citizens of the cognitive-knowledge economy.
This is fourteen distinct kinds of controller, and a quarter of a hundred when we count the sub-units. The point of writing them all down is not exhaustiveness for its own sake. The point is that each of them issues control actions — directives, allocations, investments, releases, mandates, statutes, lawsuits, salaries — and each of them receives feedback — financial returns, benchmarks, statistics, news, lawsuits, elections, market prices. The shape and the rung of those flows is what determines whether the system as a whole works. A strategic essay that talks about "the AI industry" or "the labour market" without pulling these controllers apart cannot see, for example, that the antitrust authority's process model treats concentration as something that happens in mergers, while concentration in this market is happening in long-duration compute-allocation contracts that no merger review touches.
We can summarise the structure in a single picture. Money flows downward — capital allocates to laboratories and compute providers; laboratories pay compute providers for training; compute providers pay foundries for chips; everyone hires people. Capability flows outward — laboratories release models, application companies wrap those models into products, products reach workers and consumers. Information flows in two directions — capability claims and incident reports flow upward from labs to capital, regulators, and the public; benchmarks and prices flow horizontally between labs and the open-weights ecosystem; employment and price statistics flow upward from the labour and consumer base to ministries and to the polity. State controllers issue rules and enforcement actions downward; backlash flows upward when other channels fail. The picture is dense, but it is not infinite. In the analysis companion to this chapter, every arrow is enumerated; here we only need the shape.
The diagram below shows the principal controllers and the channels between them, with rung tags on the arrows and the five mismatches (M1–M5) marked at the boundaries where they occur. Solid arrows are control actions; dashed arrows are feedback.
When we tag the arrows with their justificatory rungs, five specific mismatches stand out. Each of them is a particular case of one of the dangerous patterns we introduced in the previous section. Each is a structural reason why specific failures occur.
3.1 The first mismatch: lab safety frameworks claim more legitimacy than they operate¶
Frontier laboratories publish documents — responsible scaling policies, deployment policies, constitutional AI charters — that claim deliberative legitimacy. Their tone is the tone of public ethics: the lab is binding itself, on behalf of humanity, to thresholds that cannot be crossed without specified cause. The language sits at rung 6, the rung at which questions about what should be done are settled.
The operating mechanism, however, is rung 1. The lab is bound because the lab has chosen to be bound. The lab amends its framework when its own governance decides to amend it. There is no external party with standing to enforce, no court to which violations can be brought, no parliament that can refuse a budget on a non-compliant lab. This is the claim/operation gap. The gap is not a sign of bad faith — many of these documents are written in good faith by people who believe in them. The gap is structural: a self-binding commitment, however sincere, is operationally a rung-1 commitment, regardless of how it sounds.
The damage is that public trust calibrates to the higher claim and gets eroded at the rate of the gap. Every release that rolls back a previous threshold, every quiet amendment to an evaluation protocol, every framework that survives commercial pressure only in form, contributes to a slow loss of legitimacy that the labs cannot themselves staunch — because the legitimacy was never theirs to issue.
3.2 The second mismatch: capital cannot evaluate the laboratories it funds¶
Frontier laboratories are funded by capital allocators — venture funds, growth equity, public markets, sovereign wealth. The funds flow downward at rung 3: financial diligence, term-sheet evaluation, comparable-round analysis. The capability evidence flows back upward through the laboratories' own benchmark selection, the laboratories' own model cards, the laboratories' own mission narrative. This is rung 1 in disguise — capability claims wrapped in the lab's chosen frame.
The result is what the engineering literature calls an asymmetric loop. The arrows exist in both directions, but the upward arrow is filtered by the framing of the downward arrow. Capital cannot ask, with the rigour its own diligence claims, whether a lab's mission claims are falsifiable: the very benchmarks against which the question would be asked were chosen by the entity making the claims. The fund can either underwrite the mission narrative as given or refuse to fund — and at the scale of capital required for frontier training runs, refusing means watching a competitor fund the same lab on more pliable terms.
The consequence is that capital structure exerts a quiet but relentless pressure on operating practice. Funding rounds occur on a quarterly cadence; safety frameworks are amended on a yearly cadence; capability releases happen between rounds. When the cadences disagree, the cadence with the money wins.
3.3 The third mismatch: the labour-policy cycle cannot keep up with the capability-deployment cycle¶
Labour ministries and education systems track unemployment, sectoral employment, vacancies, and wage statistics. These indicators arrive at rung 3 or rung 4 — empirical aggregates, sometimes cumulative — but they arrive on a quarterly or annual cycle. Retraining programmes and curricular updates respond on a multi-year cycle. Capability deployment, by contrast, moves on a months cycle: a new model is released, application companies integrate it, automation is in production within weeks.
The mismatch in cycle times means that the corrective signal — a labour-market indicator showing that a class of workers is being displaced — arrives after the cohort it would have informed is already displaced. The signal cannot land in time to affect the decisions that produced the displacement. When this happens repeatedly, the rung-3/4 channel loses its informational value, and the only channel left is rung 0: political backlash, which arrives at the speed of news cycles and elections rather than at the speed of statistics. The corrective intent of the statistical channel is overrun by the rung-0 escape valve.
This is not a story about regulators being slow in some abstract sense. It is a specific identification of two cycles whose periods no longer line up — and the structural fix is not to be smarter, but to install instruments whose triggers fire automatically on a fast index, removing the requirement that a slow channel confirm.
3.4 The fourth mismatch: incidents inside the labs do not reach outside observers¶
When something goes wrong inside a frontier lab — a misuse case caught in red-team testing, an unexpected capability surfaced late in evaluation, a deployment incident in the wild — the information first exists at rung 3, inside the lab, in the internal logs. Outside observers — the AI Safety Institute, the EU AI Office, the press, the public — depend on what the lab chooses to disclose, which arrives wrapped in rung-1 framing (the lab's authority on the meaning of its own incidents).
The loop appears closed. Labs do publish model cards and post- mortems. But the closure is in form only. The rung-3 evidence that exists inside is filtered through rung-1 disclosure before reaching the controllers whose process models it would correct. This is the same asymmetric pattern as the capital/lab loop, in a different boundary. It is the structural reason why the press keeps discovering things about labs that the labs already knew.
3.5 The fifth mismatch: antitrust looks at mergers; concentration happens in contracts¶
Antitrust authorities — the FTC and DoJ in the United States, DG-COMP in the European Union, SAMR in China — evaluate market concentration through merger review and conduct cases. Their process models track mergers and acquisitions, dominant-firm behaviour, and pricing patterns. The cycle on which they operate is years. The threshold above which a transaction triggers review is set in statute.
Concentration in this market, however, does not flow primarily through mergers. It flows through long-duration exclusive compute-allocation contracts between hyperscale providers and frontier laboratories, through bundled tying arrangements between chip designers and their largest buyers, through grandfathered grid-interconnection rights at data-centre sites. None of these trigger merger review, because none of them are mergers. They happen at the contract level, on a quarters cycle, and they foreclose entry by smaller actors before any antitrust authority has cause to look.
This is cross-loop rung imposition in the third dangerous- patterns sense. The rung-2 forum (statutory antitrust review) is being asked to adjudicate a dispute generated below its threshold and beneath its cadence, and it cannot. The forum is the wrong shape for the adjudication. The structural fix is not to make antitrust faster; it is to give the forum a different statutory mandate that captures contract-level conduct.
3.6 What having five mismatches in hand changes¶
A strategic essay can speak generally about "the AI industry's governance problem." Naming five specific mismatches changes the conversation in three ways.
It changes what is observable. Once we know that the lab/capital loop is asymmetric, we can look for the symptom: capability claims that survive funding rounds but not independent evaluation. We can know what would falsify the diagnosis, and what would confirm it.
It changes what counts as a remedy. A general call for "better AI governance" is not a remedy. A specific structural intervention that turns an asymmetric loop into a symmetric one — say, by inserting a statutory pre-deployment evaluation channel that does not pass through the lab's own benchmark selection — is a remedy. The intervention does not depend on anyone changing their mind; it changes the architecture.
It changes what survives political turbulence. Personal commitments and voluntary frameworks last as long as the people and the incentives that sustain them. Structural changes — new statutes, new institutions, new contractual rules — last as long as the political settlement that created them, which is typically much longer. The remedies the engineering literature catalogues are the structural kind.
Key Finding — the five rung mismatches
Tagging every flow in the AI control structure with its justificatory rung surfaces five specific structural mismatches:
- M1. Lab safety frameworks claim rung-6 deliberative legitimacy but operate at rung-1 self-binding (Pattern B — claimed-rung inflation).
- M2. Capital flows downward at rung-3 due-diligence; capability evidence flows back at rung-1 lab-selected benchmarks (Pattern A — asymmetric loop).
- M3. Labour-policy cycle (years) is mismatched to capability-deployment cycle (months); the rung-3/4 corrective channel arrives after the cohort it would have helped is already displaced.
- M4. Incident evidence exists at rung-3 inside labs but reaches outside observers wrapped in rung-1 disclosure (Pattern A again, at a different boundary).
- M5. Antitrust operates at rung-2 with a years cycle and a merger-review threshold; concentration happens at rung-3 in long-duration contracts on a quarters cycle and below the threshold (Pattern C — cross-loop rung imposition).
These five mismatches account for most of the loss scenarios developed in §4 and motivate most of the remedies in §5.
We turn now to what specifically goes wrong when these mismatches are not addressed.
4. The nine ways the system can fail¶
A loss scenario, in the engineering vocabulary, is a description of how a particular bad outcome arises from particular failures inside the structure. It names the controllers whose process models go wrong, the feedback channels that go missing or get filtered, the controlled processes that drift into states from which exit is costly, and the external disturbances that finish the job. A loss scenario is, in effect, a small story about how the system produces harm, told in the language of its own architecture.
The full analysis identifies nine such scenarios for the AI system. They are not predictions. Each of them describes a pathway by which a specific identified harm could be produced, traceable to specific mechanisms one could intervene on. Some pathways are already partly running; others are latent. None is inevitable. The analytic purpose of writing them down is to attach each remedy to a specific failure pathway, so that the policy debate has a structural rather than a rhetorical anchor.
4.1 Concentration entrenches at the infrastructure layer¶
The first scenario is about how AI's economic gains end up concentrated in the hands of a small number of firms — but not through the obvious channel one would expect.
Antitrust authorities, as we saw in section 3.5, look at mergers. They look in the wrong place. The actual mechanism of concentration is contractual. A hyperscale cloud provider signs a multi-year exclusive capacity agreement with a frontier laboratory. A chip designer commits its leading-edge supply, a year ahead, to its three largest buyers. A foundry pre-allocates capacity windows. A data centre receives a grid-interconnection right while smaller applicants sit in a queue. Each of these is a single transaction that, on its own, looks like ordinary commerce. None of them crosses a merger threshold. None of them appears in any one controller's process model as concentration.
The cumulative effect is foreclosure. Smaller laboratories, open-weights publishers, application companies that depend on non-frontier compute, all find that the affordable supply has been contracted away in advance. Entry becomes a financial impossibility. The controlled process — the structure of the AI infrastructure market — moves to a state from which exit is costly: the contracts have already been signed; the grid interconnections have already been allocated; the leading-edge fab capacity is forward-booked for years.
The corrective channel exists, in principle. Concentration ratios, prices, and entry rates are measured. But they are measured at the slow rate at which antitrust evidence accumulates — over years — while the foreclosing transactions happen quarterly. By the time the evidence is strong enough to act on, the architecture is already locked in. The remedy class is cadence-matched controls: rules that fire on the transaction cycle, not on the evidence cycle. The specific remedies in the catalogue are reservation requirements on leading-edge supply (a defined fraction must be sold to non-frontier customers), an expedited review window on long-duration contracts during defined capability-jump moments, and a merger-review reform that captures contract-level foreclosure in the same statutory net as ordinary M&A.
4.2 Lab safety frameworks become ceremonial¶
The second scenario follows directly from the first and second mismatches, working together. Capital pressures laboratories to release on a fast cadence to defend market share. The laboratories' safety frameworks claim deliberative legitimacy. The frameworks are amended between releases, in response to commercial pressure, to allow the next release. The amendment process is internal; the legitimacy claim is external. The gap widens each cycle.
The damage here is not principally that any particular release is unsafe. It is that the legitimacy claim erodes. Each visible amendment to a safety threshold, each evaluation that produces a result the lab finds inconvenient and reframes, each model that ships with caveats no external party could have written, drains the rung-6 reservoir the labs implicitly drew on when they published their first safety frameworks. At some point the public notices that the frameworks are advisory, the amendments are unilateral, and the claims are aspirational. When that happens, the labs can no longer issue legitimate safety assurances of any kind.
The remedy class here works in two directions at once. It lowers the claim: the lab admits, structurally, that its frameworks are self-binding rung-1 commitments — useful, but limited. It raises the operation: an external evaluator with statutory authority enters the loop. The specific remedies in the catalogue are mandatory pre-deployment evaluation by an independent body (the AI Safety Institutes are early examples, operating today as voluntary partners but candidates for statutory upgrade), structural separation of safety governance from commercial governance inside labs, and a capital-structure mission-constraint mechanism that gives a public agent — a state attorney general, a regulator — standing to enforce the mission constraint, rather than relying on shareholder litigation that will never come.
4.3 Workers cannot keep up with capability deployment¶
The third scenario is about the labour market and runs through the third mismatch. Capability deploys on a months cycle — model release, application integration, deployment in customer workflow. Labour ministries' instruments — retraining programmes, curriculum updates, unemployment-insurance schemes — operate on a years cycle. Cohort-specific damage to early-career cognitive workers — the hardest-hit group because their tasks are most substitutable — is invisible in the headline employment statistics that ministries track.
The corrective channel is structurally late. By the time the employment data shows a problem, the cohort the data described has already entered the labour force, found that the entry-level roles they trained for are gone, and either dropped out, accepted underemployment, or organised politically. The political channel operates at rung 0 (electoral backlash) on a fast timescale that the rung-3 statistical channel cannot match. The rung-0 channel arrives, in other words, exactly when the rung-3 channel was supposed to have prevented its arrival.
The remedy class is automatic stabilisers indexed to a fast benchmark of task displacement, not to headline employment. Indexed unemployment insurance, indexed earned-income tax credit, indexed retraining credits: instruments whose triggers fire without requiring legislative confirmation per cycle. To these add cohort-specific transition programmes for the early-career group — recognising that the affected cohort changes year by year as substitution sweeps through different occupations — and portable benefits that decouple healthcare and pension continuity from continuous employment with a single employer. None of these require new theory; they require the index, the statutory trigger, and the political decision to commit ahead of the displacement rather than after.
4.4 The productivity surplus accrues to capital and stays there¶
The fourth scenario is the distributional one. Suppose the productivity gain from AI is real — every careful estimate suggests it is, on the order of fractions of a percentage point per year of GDP growth, possibly more. The question is who receives it.
The mechanisms by which a productivity gain reaches workers, as wages, depend on labour bargaining power. The mechanisms by which it reaches consumers, as lower prices, depend on competitive markets. The mechanisms by which it reaches the public, as taxation, depend on a fiscal architecture calibrated to the new factor shares. None of these is automatic. If labour's bargaining power falls (because AI substitutes for substitutable cognitive tasks), the wage channel narrows. If markets concentrate (scenario 4.1), the price channel narrows. If tax instruments calibrated to corporate income do not adjust for the shift toward AI-as-capital that flows to less-taxed bases, the public channel narrows.
The structural diagnosis is that no controller in the existing system is positioned to act on cross-sectoral productivity-gain capture as such. Tax authorities track corporate income, capital gains, and payroll — three categories that together miss the factor-share shift. There is, in the engineering vocabulary, a controller absent from the loop. The remedy class is to add one: a transmission mechanism, of which the catalogue lists four candidates — a productivity-indexed dividend (analogue: Alaska Permanent Fund, but indexed to AI productivity rather than oil), a value-added tax on AI inference with hypothecated dividend, an expanded earned-income tax credit indexed to the displacement benchmark, and sovereign-wealth participation in AI infrastructure with public dividend.
These are not interchangeable. The productivity-indexed dividend is conceptually clean but politically the hardest to land. The inference VAT is operationally simpler because it taxes the visible economic flow. The EITC expansion is the most familiar because the instrument already exists. Sovereign-wealth participation requires a fund and a governance mechanism; some jurisdictions have one (Norway, Singapore, the Gulf), the United States does not.
The point is not which one is correct; it is that without at least one of them, the productivity surplus has no transmission channel to labour, consumers, or the public. It will accrue to capital not because capital is rapacious but because capital is the only controller in the loop with a defined claim on it.
4.5 The open-weights price-setting mechanism collapses¶
The fifth scenario concerns a controller most strategic essays do not name as a controller at all: the open-weights ecosystem. Chinese laboratories and a handful of Western publishers release model weights free for download. The releases serve many purposes, but at the system level they perform one specific function: they set a price ceiling on the model layer. Frontier laboratories cannot price closed-weight inference far above what an adequately-equipped competitor can do with a recent open-weight release. The ceiling exists.
If the ceiling collapses, the model layer becomes a place where a small number of closed-weight providers can extract whatever the market will bear. The collapse can happen in three ways. Western governments could regulate open-weights releases out of existence — well-intentioned, on misuse-risk grounds, without recognising that the regulation also removes the price-setting mechanism. Capital could flee the few Western open-weights labs whose business is harder than their closed-weights competitors', leaving Chinese labs as the only credible open-weights suppliers — at which point the political acceptability of using their weights becomes a separate concern. Or Chinese labs could shift strategy and stop releasing open weights, for reasons of their own, and no one outside has the leverage to insist.
The diagnostic point is that the open-weights ecosystem performs a function that no controller is assigned to protect. Regulators treat open weights primarily as a misuse risk; market participants treat them as a competitive threat or competitive crutch. No body publishes a measure of the frontier-versus-open-weights gap as a system-level state variable that regulators are obliged to consider. The remedy class is what the dangerous-patterns catalogue calls domain-appropriate rung containment: route open-weights questions to a forum that holds both misuse risk and price-setting in superposition, and require any rule affecting open weights to document its impact on the gap. Concretely: publish the gap, calibrate misuse-risk rules narrowly enough to preserve the ecosystem below those thresholds, and treat the ≤12-month gap as an explicit policy goal rather than an emergent outcome.
4.6 Compute competition escalates to kinetic conflict¶
The sixth scenario is the most extreme. Compute is geographically concentrated. Leading-edge logic chips come overwhelmingly from a single foundry on a single island. Hyperscale capacity is concentrated across a small number of operators in a small number of regions. Energy and grid interconnections are concentrated where electricity is cheap. The geography of AI infrastructure is fragile in ways the geography of, say, the global oil supply once was — a small number of physical chokepoints whose disruption would stop the flow.
National security apparatuses on both sides of the US–PRC relationship watch each other's compute build-out. Each side's process model treats the other's build-out as preparation for adversarial use. The defensive interpretation — each side is building because each side fears being shut out — is filtered out on both sides by the rung-1 strategic identity that does not admit the other's defensive logic. There is no confidence-building channel at the AI-infrastructure level. The only available channels are coercive (export controls, sanctions, designations) and rhetorical (state speeches).
A scenario in which a regional crisis, perhaps unrelated to AI itself, transmits directly into the AI-infrastructure layer is not implausible. A blockade or a strike that disrupts foundry output. A sanctions package that pulls one side into commercial escalation. A reciprocal designation that classifies a category of research on both sides simultaneously. Each step is small; the cumulative path is from rung-2 rule-making through rung-0 coercion toward kinetic action.
The remedy class here is the hardest in the catalogue and the most important. An independent inspection or observation regime for AI-infrastructure deployments — analogous to the IAEA for nuclear material, the OPCW for chemical weapons, or the now- defunct Open Skies regime — would provide a rung-3 verification channel that bypasses the rung-1 filter on both sides simultaneously. It would not solve the conflict; it would create a forum in which evidence about the system's actual state could enter both sides' decision-making, reducing the chance that either side acts on a process-model error. The argument for such a regime, in the engineering vocabulary, is structural rather than rhetorical: the conflict is being driven by mutual process-model error, and only a r3 channel external to both sides can correct mutual error symmetrically.
This will not be built quickly. It depends on a precondition — shared technical vocabulary about what AI-infrastructure verification could even mean — that current track-1.5 dialogues are only beginning to produce. It depends, further, on geographic diversification of fab capacity, so that verification has somewhere to look that is not already monopolised. The sequencing is therefore: track-1.5 dialogue first (months), fab diversification in parallel (years), inspection regime later (decade).
4.7 The Chinese mirror¶
The seventh scenario is a specific consequence of the second, mirrored across the bloc boundary. Inside the Chinese state-lab system, the same Pattern A asymmetric loop operates that we identified in section 3.2 between US capital and US labs. The party centre receives capability evidence from PRC labs filtered through r1 mission-conformant disclosure; PRC labs cannot, in their own internal process, fully separate r3 evidence from r1 authority because the r1 structure is constitutive rather than contingent.
The damage is symmetric to what happens in the US case, with one critical difference. The US side is open enough that external r3 channels (independent benchmarks, AISI evaluations, academic critique) can in principle reach the apex. Whether they do depends on whether structural separation is enforceable. The Chinese side has no equivalent external channel: the structure that produces the asymmetry is the same structure that excludes external correction.
The implication is uncomfortable but operationally important. Unilateral remedies on the US side can produce evidence that PRC analysts read; capability claims that survive Western r3 evaluation also survive their own internal r1 wrapping. An externalised r3 channel becomes a correction channel for both sides at once. But the only remedy that works directly on the PRC side is a bilateral one — an inspection regime of the kind described in scenario 4.6, which exists outside both national hierarchies. This is one of several reasons that the bilateral remedies are higher-leverage than they appear.
4.8 The information environment degrades faster than authentication scales¶
The eighth scenario is about epistemics. The volume of AI- generated content already exceeds the capacity of human readers to authenticate it manually and is approaching the capacity of automated detection systems to track it. Provenance signals exist — cryptographic content credentials, watermarking — but their adoption is voluntary, and adversarial actors strip them at low cost. Information platforms whose business model is engagement-driven amplification have no incentive to favour verified content over high-engagement AI content; moderation, if it happens, imposes a unilateral cost on a single platform in a multi-platform competition for attention.
The structural diagnosis is that the controlled process — the information environment — is moving to a state in which trust in any signal degrades trust in all signals, and from that state recovery is difficult. The corrective feedback channel (measurement of provenance failure rates) does not exist at the aggregate level; only individual incidents are tracked. The controller responsible (information regulators) tracks takedowns, not aggregate state.
The remedy class is to move provenance from the platform layer (where it can be stripped) to the capture layer (camera, OS) and the transmission layer (model output API, where it can be cryptographically attested). To this add a statutory regime for aggregate measurement of provenance-failure rate, so that the question "is the information environment getting worse, by how much" can be answered. Watermarking alone is insufficient; the analogy is with seatbelt laws — voluntary mechanisms have a ceiling; mandatory mechanisms with measurement do not.
4.9 Rules calibrated to capability tiers already obsolete¶
The ninth scenario is the regulatory cadence one, mirrored. Legislatures pass AI rules calibrated to a defined capability threshold — typically a specific compute or capability level named in the statute. Algorithmic-efficiency improvements move the threshold under the rule by a factor of three or four every year. The regulator's process model treats the threshold as a fixed boundary; in reality it is a moving target. Five years after the rule passes, the rule applies to capabilities the intended scope did not anticipate, while capabilities that match the intended scope have moved into a less-regulated regime.
The corrective mechanism — recalibration — depends on re-passage of primary legislation, which requires legislative attention that is itself rationed. By the time the rule is re-passed, the threshold has moved again. The remedy class is to write rules whose thresholds move with a published benchmark, and to grant a standing body the authority to recalibrate without primary legislation, subject to a notice-and-comment process. This is not novel — securities regulators in many jurisdictions recalibrate margin and capital requirements without primary legislation — but it requires a deliberate political decision to delegate, and the political appetite for delegation varies by jurisdiction.
4.10 What having nine scenarios gives us¶
Nine scenarios is more than a strategic essay would provide and fewer than the underlying analysis admits. The number is roughly right because each scenario corresponds to one identifiable failure mechanism. Together they cover all the unsafe control actions catalogued in the analysis; each remedy class points at specific UCAs whose failure pathway it interrupts.
The nine scenarios are not independent. Scenario 4.1 (concentration) makes scenario 4.4 (surplus capture) worse. Scenario 4.2 (ceremonial safety) makes scenario 4.8 (epistemic degradation) worse. Scenarios 4.6 and 4.7 (the geopolitical pair) interact symmetrically with each other. The remedies, in turn, are not independent: as we will see in the next section, several of them share preconditions, several form sequencing chains, and the structurally hardest remedies are exactly those whose absence would cost the most to live with.
5. What can be done¶
Twenty-two specific remedies follow from the analysis. They divide into five families, each addressed at one or more of the nine loss scenarios. The family-level summary first; then the sequencing.
The first family is measurement and feedback channels. Several of the most damaging mismatches occur because the system lacks an honest measurement of its own state. Regulators do not know the frontier-versus-open-weights gap; nobody publishes the aggregate provenance-failure rate; labour ministries track headline employment but not task-by-task displacement. Four remedies fill these channels: a quarterly assessment of the open-weights gap, a quarterly aggregate provenance-failure measurement, a task- displacement benchmark indexed by occupation cohort, and a track-1.5 dialogue process that builds shared technical vocabulary between the US and PRC technical communities about what AI-infrastructure verification could mean. Each of these is operationally close to existing instruments and politically modest. Each of them, however, is a precondition for remedies in later families.
The second family is rules built on the new measurements. Once the open-weights gap is published, a targeted misuse-risk regime can be calibrated narrowly enough to preserve the ecosystem below those thresholds. Once the provenance-failure rate is measured, mandatory provenance becomes specific rather than gestural. Once the task-displacement benchmark exists, automatic stabilisers — unemployment insurance, retraining credits, EITC top-ups — can be indexed to it, so that triggers fire on the fast cycle that the disturbance arrives on. Reservation requirements on leading-edge supply, capability-tier rules indexed to a moving benchmark with sunset clauses, and a standing recalibration authority for AI rules each address a specific identified pathway. A mandatory pre-deployment evaluation regime — turning the AI Safety Institutes from voluntary partners into statutory ones — closes the asymmetric loop between labs and capital.
The third family is structural remedies. These are the interventions that change the architecture rather than the operating parameters. Merger-review reform that captures contract-level foreclosure. Structural separation of safety governance from commercial governance inside frontier laboratories. A statutory public-benefit-corporation class with mission constraints enforceable by a public agent, not by shareholder litigation. Cohort-specific transition programmes for early-career cognitive workers. Portable benefits that decouple healthcare and pension continuity from continuous employment. A value-added tax on inference, hypothecated to a public dividend. Sovereign-wealth participation in AI infrastructure, where the jurisdiction has a fund. These are harder to land politically and slower to take effect, but each of them addresses a structural failure that no operating- parameter adjustment can fix.
The fourth family is the slow geopolitical remedies. Three of the most consequential remedies operate on a decade horizon because their preconditions are physical or diplomatic. Geographic diversification of fab capacity — bringing leading- edge logic-chip production to a strategically meaningful share outside the present chokepoint — is funded but slow; the constraint is sustained political commitment across electoral cycles. The independent inspection or observation regime for AI infrastructure depends on shared technical vocabulary, which the track-1.5 process slowly produces, and on geographic diversification, so that verification has somewhere to look that is not already monopolised. A productivity-indexed dividend, of the Alaska-Permanent-Fund kind but indexed to AI productivity, operates on a decade horizon because it requires both a fiscal instrument (which the inference VAT could provide) and a governance mechanism (which sovereign wealth could provide). Each of these is the long-cycle remedy whose absence the others cannot make up for.
The fifth family is the politically novel ones. Two remedies do not fit cleanly into any of the four prior categories. An expedited capability-jump merger-review window, granting antitrust authorities a temporary hold on long-duration compute contracts during defined capability-jump moments, is a small intervention with a precise function but no real precedent. An employer-independent benefits regime is operationally familiar but politically contested in jurisdictions where employer-tied benefits define the post-war social contract. These are listed separately not because they are less important, but because they require political adoption on terms the system does not yet provide.
5.1 The four-phase sequence¶
The 22 remedies do not all enter the world at once. They enter in a specific order constrained by the dependency graph among them, and the analysis derives a four-phase implementation plan.
The first phase, months one through eighteen, is measurement. The four feedback-channel remedies — the open-weights gap, the provenance-failure rate, the task-displacement benchmark, the track-1.5 dialogue — can begin without statutory change. An AI Safety Institute could initiate the open-weights gap publication tomorrow, as a non-binding research output. A national statistics office could establish a provenance-measurement programme on existing authority. A labour-statistics office could publish a task-displacement index using existing labour data and AI task mapping. The track-1.5 dialogue is already running in several forums; sustaining and broadening it requires nothing more than philanthropic funding and continued attention. By month twelve, each of these channels will have produced a first batch of evidence; by month eighteen, the question of whether to formalise each one statutorily will have a real answer.
The second phase, months twelve through forty-eight, builds rules on the new feedback channels. This is where most of the operationally familiar remedies sit: indexed automatic stabilisers, indexed EITC, mandatory provenance, reservation requirements on leading-edge supply, capability-tier rules with sunset clauses, a recalibration authority, mandatory pre- deployment evaluation, a targeted misuse-risk regime for open weights. Each of these is a piece of legislation or a regulation written against a measurement that the first phase produced. Each is operationally close to existing instruments. Each requires political work but not novel theory.
The third phase, months thirty-six through eighty-four, is the structural remedy phase. Merger-review reform, structural separation of lab safety and commercial governance, the statutory PBC class with public-agent enforcement, cohort-specific transitions, portable benefits, the inference VAT with hypothecated dividend, sovereign-wealth participation. These require political capital that the first two phases will have generated, and an empirical record that shows whether the operating-parameter remedies were sufficient on their own (the analysis suggests they will not be, but the evidence by month thirty-six will be specific). Each of these is structurally novel and politically harder than the second-phase remedies; each addresses a pathway the second-phase remedies cannot close.
The fourth phase, months sixty through one hundred and forty-four, is the long-cycle remedy phase. Fab diversification will already be running in parallel; what completes in this phase is the inspection regime, the productivity-indexed dividend, and the expedited capability-jump review window if it has not already been adopted. These are the remedies whose absence imposes the largest tail cost — concentration unbroken, geopolitical escalation unverified, surplus capture untransmitted — and whose adoption is hardest because their political horizon exceeds electoral cycles.
The decision-memo companion to this chapter sets out the dependency graph in full and identifies, at each phase boundary, the question that the prior phase must have answered before the next phase becomes feasible. By month twelve: did anyone publish the open-weights gap? By month eighteen: is the displacement benchmark stable? By month thirty-six: have at least three of the eight phase-two rules been adopted? At each decision point, the analysis describes both what success looks like and what failure implies for downstream sequencing. This is what distinguishes a planning artefact from a wish list.
The four-phase sequence at a glance
| Phase | Months | Theme | Representative remedies |
|---|---|---|---|
| 1 | 1–18 | Measurement and feedback channels | Open-weights gap publication · provenance-failure measurement · task-displacement benchmark · track-1.5 dialogue |
| 2 | 12–48 | Rules built on phase-1 measurements | Mandatory pre-deployment evaluation · indexed automatic stabilisers · indexed EITC · mandatory provenance · reservation requirements · capability-tier rules with sunset · recalibration authority · targeted misuse-risk regime |
| 3 | 36–84 | Structural remedies | Merger-review reform for contract foreclosure · structural separation of safety governance · statutory PBC class · cohort-specific transitions · portable benefits · inference VAT · sovereign-wealth participation |
| 4 | 60–144 | Long-cycle remedies | Geographic fab diversification · independent inspection regime · productivity-indexed dividend · expedited capability-jump review |
Each phase's gating decision determines whether the next phase is structurally feasible. Phase 1 is reversible and operationally familiar; phase 4 is largely irreversible once started.
5.2 The non-state actor's leverage¶
The remedies are not, as a class, things that one private actor can do alone. They are political and statutory. But several of them have entry points where a non-state actor — a foundation, a research institute, a coalition of application companies, a group of labour organisations — can move first and, by moving first, make the statutory work tractable.
Phase-one measurement is the cleanest example. Each of the four channels can be initiated as a non-binding publication; once the data exists, the statutory mandate becomes specific rather than gestural. Drafting model-bill language for phase-two rules is work that legal-policy institutes do well, and adoption rates of model-bill language are unusually high in jurisdictions where the legislature is overstretched. Coalition-building among open-weights-dependent companies is work that market participants can do without waiting for political alignment. Track-1.5 dialogue is the canonical philanthropic-funded geopolitical intervention; it is not a substitute for state action, but it produces the vocabulary state action requires.
The leverage point for any non-state actor, then, is to ask which of the 22 remedies in which phase has an entry point that matches the actor's position, and to commit to the entry-point work for at least one full decision cycle. The dependency graph makes the leverage points explicit; it also makes clear which work, however well-meaning, is not on the critical path.
5.3 Reversibility¶
A final consideration constrains the remedy set. Most of the proposals are reversible at moderate cost within their first cycle. Statistical publications can be discontinued. Indexed triggers can be recalibrated. Pre-deployment evaluation requirements can be relaxed. Reservation requirements can be adjusted. This is structurally important because it means a risk-averse stakeholder can support phase-one and most of phase-two with limited downside. The risk of getting the rule wrong is not the risk of locking in an error.
The exceptions are the structural and slow-cycle remedies. Once established, an inspection regime is hard to withdraw because withdrawal would itself signal an escalation posture. Sovereign- wealth participation in AI infrastructure cannot be unwound without selling at unfavourable timing. Fab diversification is sunk capital. A productivity-indexed dividend, once it becomes a household income source, is politically irreversible in practice. A loss-averse stakeholder considering tail risks — the kinetic- escalation risk in scenario 4.6, the cohort-displacement risk in scenario 4.3 — would weight these irreversible remedies more heavily despite their cost. A risk-averse stakeholder weights phase-one and phase-two preferentially.
The point of putting reversibility in the analysis explicitly is that different stakeholders, with different risk tolerances, will rank the same remedies differently. The analysis does not adjudicate the ranking. It surfaces the trade-off so that the adjudication can happen on terms a deliberative process can accept.
6. What the analysis does not do¶
A method that produces twenty-two specific remedies is also a method that risks being read as a forecast or a manifesto. It is neither, and it is worth being explicit about what the limits are before closing.
It does not predict which of the nine loss scenarios will occur, or in what combination, or on what timescale. The scenarios are possibility statements: each one describes a pathway from the current structure to a specific bad outcome, and the structure does in fact have those pathways. Whether the outcomes occur depends on which control actions are issued, by which controllers, in which sequence, and with what feedback. The analysis maps the terrain; it does not tell the reader how the journey ends.
It does not settle value disputes. The remedies are derived from the loss scenarios, and the loss scenarios are derived from stakeholder-named losses. Different stakeholders would name different losses. A reader who does not consider concentrated ownership of the AI substrate to be a loss — perhaps because they believe the firms in question are well-governed and well- intentioned — will not find scenario 4.1 a problem and will not weight remedies R-1 through R-3 highly. The analysis cannot adjudicate that disagreement; it can only show that if the loss is taken to be a loss, then certain remedies follow with specific properties. The reader's own values determine which "ifs" are live.
It does not tell any particular government or firm what to do. Decisions belong to controllers; the analysis describes what each controller could in principle do and what consequences each choice would have for the controlled processes. A decision-memo addressed to a specific stakeholder — say, a foundation allocating its grant budget over five years, or a labour union prioritising its lobbying agenda, or a cabinet office choosing between policy packages — would be a different document, derived from this analysis but committing to that stakeholder's goals and constraints.
It does not, finally, exhaust the possible failure modes of the AI system. The analysis is a first cycle in the engineering sense: every loss scenario is one that the controller / control- action / feedback / process-model decomposition can surface. There may well be failure modes that this decomposition cannot see — failures that emerge from interactions across multiple loops simultaneously, failures that depend on non-controllers the analysis treats as environment, failures whose causes lie below the threshold of the current decomposition. A second cycle, with refined boundaries and a tighter decomposition, would find some of these. The analysis we have is bounded in the way every engineering analysis is bounded: by the granularity at which the system was modelled.
7. Coda: the question reformulated¶
We started with three questions: about jobs, about power, about geopolitics. The structured analysis allows us to reformulate each of them in a way that points at action.
The jobs question — will AI cause mass joblessness or freer prosperity? — is, structurally, the question of whether the labour-policy cycle (years) will be brought into alignment with the capability-deployment cycle (months) before the cohort displacement that the mismatch produces becomes irreversible. The technology produces the surplus; the institutions decide whether the surplus reaches workers, consumers, the public, or only capital. The analysis identifies the specific instruments that would close the cycle gap (indexed automatic stabilisers, a displacement benchmark, cohort-specific transitions, portable benefits) and the specific instruments that would transmit the surplus (the inference VAT, the productivity-indexed dividend, the EITC top-up, sovereign-wealth participation). Whether any of these is adopted is a political choice made under known trade-offs — not, as the strategic essay would have it, an open question with a wide range of plausible answers.
The power question — will a few firms own everything? — becomes, structurally, the question of whether the contract-level foreclosure mechanism currently moving compute infrastructure into a small number of hands is captured by an antitrust regime calibrated to the right cadence and the right threshold. The analysis identifies merger-review reform, reservation requirements on leading-edge supply, and an expedited capability- jump review window as the specific instruments. It also identifies the open-weights ecosystem as a price-setting mechanism whose suppression would be the worst single thing that could happen at the model layer, and identifies the specific regulatory pathway (well-meaning misuse-risk regulation that fails to recognise the price-setting function) by which the suppression could occur.
The geopolitics question — will the US–PRC AI competition escalate? — becomes, structurally, the question of whether a rung-3 verification channel can be installed between the two blocs before mutual process-model error converts a series of rung-2 coercive moves into a kinetic exchange. The analysis identifies track-1.5 dialogue as the precondition, fab diversification as the parallel constraint, and an infrastructure-inspection regime as the long-cycle remedy. It also identifies the symmetric Pattern-A failure on both sides that makes mutual error structurally likely, and explains why unilateral US-side remedies (especially mandatory pre-deployment evaluation) produce evidence that bypasses the rung-1 filter on the PRC side as well, making them more valuable than they appear when each is considered in isolation.
None of these reformulations resolves the questions. Each of them turns a forecasting question into a structural question, and a structural question is one a deliberative process can act on. A political community that refuses to take action on a forecasting question — because the future is uncertain, because experts disagree, because the right thing to do is unclear — can take action on a structural question, because the structural question is not "what will happen" but "what feedback channels are missing and what institutions are needed to install them." The first question admits indefinite postponement; the second does not.
That, finally, is what an engineering reading offers that an essay does not. The engineering reading produces a list of specific things that could be done, in a specific order, by specific actors, with specific costs and specific reversibility profiles. It does not promise that any of them will be done. It does promise that, if one of them is done, the system will respond in the way the analysis predicts — because the prediction is not a guess about behaviour but a derivation from structure.
Whether the structures change is, in the end, the same question as whether the political communities affected by AI choose to change them. The technology will not make the choice. Capital will not make the choice. The labs will not make the choice. The choice has to be made by the controllers whose process models the structural analysis maps and whose feedback channels the remedies repair. And the choice can only be made if the controllers see the structure clearly enough to know which levers exist.
This chapter has tried to make the structure visible. The rest is up to whoever is reading.