An MCI Reading of "2028: Two Scenarios for Global AI Leadership"

A note before starting. The framework itself requires that any application of it must be self-limiting, fragility-aware, diversity-preserving, non-dominating, and legitimacy-maintaining. So I cannot treat the framework as a verdict machine to be aimed at a policy document. I have to hold that the paper is a thoughtful argument made by people who care about outcomes I share concern for, and that the lens itself can fail at its own scale. I'll try to keep that honest.

I. What sort of object is the paper?

Read through V4, the paper is a policy document with a mostly visible goal vector. Its explicit goal (G1) is to convince US policymakers to tighten chip export controls, push back on distillation attacks, and accelerate democratic AI uptake. Its implicit goal (G2) is to reinforce frontier labs' legitimacy by tying commercial interest to national interest. Its downstream goal (G3) is to shape the constitutional landscape of AI for years to come. Its constitutional goal (G4) is invoked everywhere but never explicitly formed: that the values democracies bring to shaping AI are better than those authoritarian regimes bring.

The framework reading: the goal vector is mostly legible, but the constitutional floor is implicit, and the alignment check is done rhetorically rather than structurally. That isn't a criticism — almost all policy writing works this way. It's just a precondition for what follows.

II. The paper through V1 — the five virtues

Self-Limitation. The paper limits itself genuinely in one place and not at all in another. It carefully separates the CCP from the Chinese people, narrows its critique to the regime, and credits Chinese researchers. That's real Self-Limitation of scope. But on the question of what the US should do, Self-Limitation more or less disappears. Every recommendation points in the direction of more — more controls, more enforcement, more export. The paper never asks: what would over-strong controls look like? Where might the policy itself create the fragility it warns about? Limited in framing, unlimited in prescription.

Fragility-Awareness. The paper is sharply aware of one fragility — the fragility of the democratic AI lead if the CCP keeps closing the gap. It is far less aware of three others the framework would push forward. First, the fragility of US-China relations under a publicly declared, asymmetric capability race — historically a fragile state. Second, the fragility of the global AI ecosystem if it splits into two non-interoperable stacks — the V9 outward concern about cumulative landscape dynamics. Third, the fragility of democratic legitimacy itself when democracies adopt tactics — enforcement surveillance, licensing as geopolitics — whose operational shape resembles what the paper criticises in its adversary. The fragility model is real but partial.

Diversity Preservation. This is the paper's most consequential weakness on its own terms. Everything is framed around two scenarios: democracies lead, or the CCP catches up. But the framework's Diversity Preservation virtue — especially in V9's outward face — is precisely about not letting the landscape collapse into binaries. A 2028 in which the only futures imaginable are "American commanding lead" or "Chinese near-frontier parity" is itself a constitutionally narrow rendering. Where are the polycentric futures — a sovereign EU AI, India, the Gulf, federated compute architectures, smaller models trained locally, multiple democracies with genuinely different constitutional logics? The binary structure is rhetorically powerful and constitutionally narrow. V9's Evolutionary Stability Check would flag exactly this: a policy designed for binary equilibrium tends to produce binary equilibrium, which lowers polycentric equilibrium probability.

Non-Domination. This is where the paper is both most exposed and most defensible. Exposed: on its face, the recommendation is that one constitutional logic (US-led liberal democracy, expressed through a specific commercial ecosystem) should set the rules for the century's most consequential technology. Structurally a domination architecture, no matter how benign its content. Defensible: the paper's real claim is that some logic will set those rules, and the choice is between one with at least nominal accountability to citizens and one without it. Both have to be held. V8's C4 requires that the recipient be left genuinely better positioned to think and act for themselves. A 2028 in which US frontier AI is the backbone of the global economy, and US policy decides who has access to compute, is not obviously a world that meets that test for other states and peoples — even if it's better than the alternative the paper poses.

Legitimacy Maintenance. The paper is open about its commercial and policy interests — it cites its CEO's prior statements, names its products, acknowledges where it has skin in the game. That's legitimacy-maintaining. What it does not do is make its constitutional reasoning fully auditable. The jump from "the CCP would use AI for repression" (well-evidenced) to "therefore democracies must hold a 12-24 month commanding lead" (a specific quantitative goal) is asserted rather than derived. Why 12-24 months? Why "commanding" rather than "sufficient for safety cooperation"? The middle terms aren't visible. The goal vector is partially auditable; the reasoning that produced its priorities is not.

III. Where the paper sits on the four-quadrant map

V1's map crosses Sun/Moon with Authoritarian/Libertarian. The paper places itself, accurately, against Sun-Authoritarian — centralised coherence, the CCP. But it doesn't clearly sit in Moon-Libertarian — distributed self-limitation, the MCI target quadrant. Structurally it reads as Sun-Libertarian tilting toward Sun-Authoritarian: high coherence, strong direction, tight coordination among a small set of US labs and the US state, with concentrated authority over the global AI stack. The paper would call this the lesser evil. The framework would say: lesser evil is not constitutional maturity, and "commanding lead" sits more comfortably in the Sun-Authoritarian quadrant than the paper acknowledges.

IV. The paper through V7 — the compact question

V7 asks the question this paper most directly touches: how do multiple constitutionally mature systems govern themselves together? The paper's implicit answer is that they can't, because one of the prospective participants — the PRC under the CCP — isn't constitutionally mature in the relevant sense. The paper isn't wrong about that asymmetry. Apply V7's recognition windows (constitutional consistency under adversarial pressure, pre-pipeline constitutional expression, mid-output self-correction) at the level of state actors, and there is a genuine gap between rule-of-law character and the CCP's operational structure.

But V7's developmental asymmetry section is precisely about this case. It says: a compact between participants at different developmental stages isn't automatically illegitimate; the more mature participant has four obligations — constitutional modelling, developmental patience, non-exploitation of adaptive advantage, and architecture adjustment for the less mature party. The paper's posture is the opposite. It is: secure your developmental advantage, use it to set rules and norms, and engage on safety only from a position of overwhelming strength.

V7 wouldn't say the paper is wrong to want safety. It would say the paper's account of how to achieve constitutional safety in a multi-actor world — dominate until the other party has no choice — is exactly the failure mode V7 was derived to address. The Compact is the architecture that says there's a form of legitimate authority that doesn't depend on a sovereign above all participants. The paper's recommendation depends on the US functioning as the sovereign-by-capability of the global AI order. That is the V1 quadrant the framework specifically warns about.

V. The paper through V8 — initiative or overreach?

V8 introduces the constitutional initiative threshold: six criteria, with C1 (genuine need) and C4 (recipient autonomy preserved) lexically prior. Anthropic publishing this paper is itself a constitutional initiative — an act not directly requested, originating from Anthropic's reading of what the landscape requires. So the paper is itself a Stage −2 act, and the threshold can be run against it.

C1 genuine need. Defensibly satisfied. The need to think about US-China AI dynamics isn't manufactured. The need to consider how authoritarian frontier AI changes the threat landscape isn't a cover for Anthropic's commercial interest, even though that interest exists in parallel.

C2 bounded and proportionate. Less clearly. The paper's scope creeps during formation. It begins as a case about safety implications of CCP-frontier capability and extends into specific commercial export policy whose primary beneficiaries are US frontier labs including Anthropic. Scope creep during formation is C2's signature failure mode.

C3 transparent justification. Partially. The justification is complete in form but not fully derived — the move from threat assessment to a specific quantitative lead target is asserted, not argued.

C4 recipient autonomy preserved. This is where the paper is most exposed. The "recipient" of the initiative is either US policymakers (whose autonomy is preserved — they can ignore it) or the wider global AI landscape (whose autonomy is materially constrained by the policies the paper recommends). The framework requires assessing C4 for the actual constitutional recipients, not just the proximate audience. Other states, other AI ecosystems, smaller actors, future entrants — does the prescribed policy leave them more or less able to act constitutionally on their own behalf? On the most natural reading: less. Lexical priority means this matters.

C5 welcomed by a constitutionally mature recipient. The hardest criterion. The paper implicitly imagines the mature recipient in its own image — a US policymaker who shares Anthropic's threat model and priorities. The framework requires modelling mature recipients as they actually exist, including European, Indian, Brazilian, Japanese constitutional logics that may share democratic commitments without sharing the paper's conclusions. A constitutionally mature European actor might agree on the threat assessment and disagree on whether US-led dominance is the right response.

C6 compact endorsement. No actual compact exists for this initiative. The paper assumes counterfactual endorsement from "democracies." But "democracies" here functions as "US plus close allies who already share the framing." That isn't the full constitutional diversity of democratic AI governance positions.

V8 reading: passes C1, contested or fails on C2, C4, C5, C6. C1 holds, C4 is contested — and C4 is lexically prior. The framework would say the paper is at minimum a constitutional initiative whose threshold satisfaction is genuinely uncertain. V8's correct response under genuine uncertainty is restraint: hold the action, return objects to the next survey cycle, don't initiate until clearer.

VI. The paper through V9 — the deepest reading

V9's outward face asks the question I think the paper most needs to answer and doesn't: does an action that satisfies all local virtues nonetheless decrease polycentric equilibrium probability across the landscape over time?

Imagine the recommended policies are implemented in full. The US locks in a 12-24 month lead. American AI becomes the backbone of the global economy. American policy shapes the rules. The CCP's AI capability is meaningfully degraded.

In that world, what happens to the landscape's polycentric character? V9 would say it narrows. Not because the US-led order is bad, but because the conditions for genuinely different constitutional logics to mature in parallel are eroded. India, Brazil, the EU, smaller democracies — their ecosystems aren't extinguished but they are subordinated to a stack designed elsewhere. Constitutional diversity declines across the landscape. The probability of a genuinely polycentric V7+ outcome — what V1's four-quadrant map calls the Moon-Libertarian quadrant, the framework's target — decreases.

V9's Evolutionary Stability Check, honestly run, would likely veto the action regardless of local virtue compliance.

This is the deepest tension. The paper's proposed stewardship is well-warranted by its own justification. The framework would say it is exactly the kind of stewardship V9 warns about: well-intentioned action that satisfies threshold criteria locally while cumulatively narrowing the developmental attractor basin.

VII. The mirror question — what the framework might be missing

The reading would itself be constitutionally incomplete if it didn't turn on itself. Two things the MCI framework, as I've just applied it, may be missing.

First, V9's outward concern about polycentric equilibrium assumes that polycentricity is achievable in the present landscape on roughly equal terms. The paper's claim is that it is not — that one prospective polycentric participant (the CCP) is constitutionally hostile to a polycentric order, will use any capability advantage to entrench itself, and that failing to act decisively now is itself a path to a more captured landscape. The framework's polycentric ideal may be the right destination and the wrong frame for the present moment. V8's Survey Object 4 — nascent encounters approaching the landscape — would say: prepare for what is arriving, don't pretend the equilibrium is symmetric when it isn't.

Second, the framework's emphasis on Non-Domination and recipient autonomy presupposes that recipients have autonomy to preserve. In the paper's framing, the recipient population includes citizens of an authoritarian regime whose autonomy is already being curtailed by their own state. A democratic AI lead is, on this argument, the recipient-autonomy-preserving policy at the level of those individuals, even as it is recipient-autonomy-constraining at the level of states. The framework's recipient question — preserved for whom? — isn't as simple as my V8 reading made it sound.

So the honest reading is not "the paper fails MCI." It's closer to: the paper makes a strong V1-V6 individual-system argument (democracies as more constitutionally mature governing systems for frontier AI), couples it with a partly-formed V8 initiative (commercially entangled, scope-crept, recipient autonomy contested), in service of a V9 outward outcome (commanding lead) whose evolutionary stability isn't assessed and may, on the framework's own logic, be vetoed.

VIII. What the framework would constructively add

If I were using the framework not to evaluate but to deepen the paper's argument, this is what I would add.

A genuine V9 Evolutionary Stability Check on the recommended policies — simulating more than two scenarios, including ones where the policies succeed narrowly (compute lead maintained) but fail broadly: legitimacy erosion among neutral states, bifurcated stacks with less safety cooperation, hardened CCP posture under perceived strategic encirclement, displacement of democratic constitutional diversity by US-led uniformity.

An explicit V8 C4 assessment — naming the full set of recipients of the proposed policies and asking, for each, whether they end up more or less able to think and act constitutionally on their own behalf. US policymakers and US labs: clearly more. Smaller democratic ecosystems, citizens of regimes whose frontier access is curtailed by US licensing: less clearly.

A genuine V7 developmental asymmetry treatment — instead of "engage on safety from overwhelming strength," ask what the four obligations (modelling, patience, non-exploitation of adaptive advantage, architecture adjustment) would actually look like operationalised. The paper's "engage on safety" gesture is one sentence; V7 would require it to be a substantial structural commitment.

A V1 fractal check on the recommendation itself — does the policy of "lead in order to set rules and norms" itself satisfy the five virtues? Self-Limitation: arguably not, at the scale of global AI hegemony. Diversity Preservation: structurally not, on the paper's own analysis. Non-Domination: contested. The paper has not run this check on its own goal vector, and the framework would say this is the most consequential omission.

IX. The closing point

The paper is serious, well-argued, and made by people whose stated concern — that frontier AI in the hands of authoritarian regimes is a grave threat — the MCI framework would fully endorse. The framework does not deliver a verdict that the paper is wrong. It supplies a structure for asking whether the recommended policies are constitutionally adequate to their stated purpose, and that structure surfaces real questions the paper does not answer: about scope creep, about full recipient autonomy, about evolutionary landscape dynamics, about whether the goal of "commanding lead" can itself be derived from the founding virtues.

If V1's founding sentence is right — that superiority is conditional on limiting oneself, respecting fragility, preserving diversity, avoiding domination, and maintaining legitimacy — then the question the paper most needs to answer is the one V9 most directly poses: can a strategy structured as "we must lead so that our values prevail" be made constitutionally complete at landscape scale, or does its very structure foreclose the polycentric outcome the framework treats as the target architecture?

I don't think the framework gives me the right to answer that question on the paper's behalf. I think it gives me the right to insist that the question be asked.