Two visions of the future — and the page missing from both
This repository already runs one argument on paired primaries: two definitions of “frontier” — the laboratories’ and the law’s — read side by side until the difference itself becomes the finding. This page runs the same method on the future. Two documents now sit in the project library, and they are the two most consequential accounts of the AI decade in circulation: the sitting administration’s official roadmap, and the leading independent forecasting group’s plan. They disagree about nearly everything — pace, posture, who should hold the wheel. Read against this Act, they agree about the two things that matter, and they fall silent in exactly the same place.
The two documents, graded
Washington’s future. Winning the Race: America’s AI Action Plan (The White House, 23 July 2025; primary PDF in hand, read in full — the verification record). Official federal policy, signed out by the President’s science adviser, his AI and crypto adviser, and his national security adviser. Three pillars — accelerate innovation, build infrastructure, lead internationally — carried by roughly ninety recommended actions. Its register is the race: “The AI race is America’s to win, and this Action Plan is our roadmap to victory.”
The forecasters’ future. AI 2040 / Plan A with the AI Futures Model behind it (the AI-2027 team, 2026; ⚠ scenario material by its own declaration — “primarily a recommendation, not a prediction”). The deep reading, with the model’s arithmetic and its epistemic grades, is owned by the forecasters’ arithmetic; this page draws on it without restating it.
One document is written by the state, to win; the other by independent researchers, to survive. That is the widest provenance spread the record offers — which is what makes their agreements evidence.
Where the two futures agree — and agree with this Act
Compute is countable, locatable, and governable. The Action Plan directs Commerce to explore “leveraging new and existing location verification features on advanced AI compute to ensure that the chips are not in countries of concern” (Pillar III) — chip-level location verification as live enforcement policy. The forecasters’ plan builds its entire treaty architecture on the same physical fact, down to priced declaration thresholds and recomputation audits. When the race document and the halt document both treat compute as the thing a government can find and count, the standing objection to this Act’s compute-denominated trigger — that 10²⁶ operations is a phantom no state could administer — is being answered from both ends of the politics at once.
Evaluations are becoming law’s instrument. The Plan, verbatim: “Over time, regulators should explore the use of evaluations in their application of existing law to AI systems” (Pillar I). Existing law, applied to AI through evaluations, is this project’s third vehicle (paths to enactment) and the logic of the Act’s interim-standards design. The forecasters reach the same instrument from the other side: safety cases that must “withstand criticism from the public … government auditors, and rival AI companies.”
Incidents are expected — by both. The Plan builds a federal AI incident-response capacity in so many words: playbooks amended, standards written, response teams equipped, because “prudent planning is required to ensure that, if systems fail, the impacts to critical services or infrastructure are minimized” (Pillar II). A government does not build fly-away kits for a risk it considers hypothetical. The Act’s reporting-on-fixed-clocks architecture (SEC. 10) is the statutory form of the same expectation.
The systems are conceded to be unpredictable. The Plan again, on interpretability: technologists “often cannot explain why a model produced a specific output,” which “can make it hard to predict the behavior of any specific AI system” (Pillar I). That sentence, in the government’s own accelerationist document, is the factual predicate for ex-ante duties: where behavior cannot be predicted, due care must attach to the decisions around the system — validation, records, reporting — because nothing else is available to attach it to.
And attestation without enforcement is conceded to fail. The Plan’s biosecurity section orders screening requirements with “enforcement mechanisms for this requirement rather than relying on voluntary attestation” (Pillar III). Sit that beside the industry’s twelve voluntary frontier frameworks, none of which requires a signature (the census): Washington’s own roadmap states the principle — attestation that costs nothing binds nobody — and this Act is that principle, applied to the attesters who matter most.
The shared silence
Now the absence. The Action Plan is not short of officers: it formalizes a Chief Artificial Intelligence Officer Council, and its incident playbooks require “Chief Information Security Officers to consult with Chief AI Officers.” Every officer it names is the government’s own employee. It even shows that conditioning frontier developers is thinkable when the objective matters — federal procurement is to be limited to “frontier large language model (LLM) developers who ensure that their systems are objective” — so the instrument of imposing terms on developers is present and armed. And yet across three pillars and ninety actions, no natural person at any frontier developer is asked to sign, certify, or answer for anything.
The forecasters’ plan runs forty-seven thousand words of bans, audits, safety cases, and burden-shifting to the same silence — no liability, no certification, no named signer — with the officers themselves identified and left unreached (the forecasters’ arithmetic, § 2).
Two futures, opposite in politics, method, and mood; one identical omission. Neither the race’s roadmap nor the halt’s blueprint ever names the person. That is the page missing from both — and it is the only page this Act is.
The footnote that reads like a confession
The Plan’s permitting section proposes to expedite data-center construction by “streamlining or reducing regulations promulgated under the Clean Air Act, the Clean Water Act, the Comprehensive Environmental Response, Compensation, and Liability Act” — footnotes 27 and 28, citing 42 U.S.C. §§ 7401–7671q and §§ 9601–9675 in full. The first of those statutes is not a neutral bystander in this project’s argument: its enforcement section is where Congress codified the words “responsible corporate officer” into federal statutory text — 42 U.S.C. § 7413(c)(6), added 1977, two years after Park (comparative officer liability, § 5; the table of authorities) — and it is the statute under which a Volkswagen engineer served federal time for software built to behave differently on the test rig (the same conduct).
This page takes no view on permitting reform. The point is narrower and better: the Action Plan’s own footnotes locate the enforcement architecture this Act extends — in-force statutes, officer-reaching, prosecution-tested — and propose relief from their paperwork for the buildings AI will run in, while proposing nothing at all about their doctrine for the people AI answers to. The infrastructure of the frontier is to be freed from the very statutes whose personal-liability spine this project asks a state to extend to it. A sponsor could open with that sentence.
What each future does to the Act’s prospects — both directions, honestly
Headwinds, recorded. The Plan would have federal agencies “limit funding if the state’s AI regulatory regimes may hinder the effectiveness of that funding” — the funding lever of the ceiling weather the half-statute page tracks — and orders FTC actions reviewed so they do not “advance theories of liability that unduly burden AI innovation.” An anti-liability instinct at the top of the federal executive is a headwind for any liability statute, and this project records it rather than rounding it away. The forecasters, for their part, have no state lane at all (the forecasters’ arithmetic, § 6).
And the lane left open — by the Plan itself. The same paragraph that arms the funding lever states that the federal government “should also not interfere with states’ rights to pass prudent laws that are not unduly restrictive to innovation.” That is the Action Plan’s own version of the reservation this record now holds four times over — the moratorium hearing’s general-applicability concession, the SANDBOX framing, the Blackburn bill’s express preservation, and now the roadmap itself. A statute that creates no agency, licenses nothing, restrains no training run, and attaches duties to persons the way food-and-drug law has since 1943 has a serious claim to being exactly the “prudent law” the sentence reserves — and a state officer amendment regulates nothing the Plan’s ninety federal actions touch.
For a reviewer or a sponsor, the yield is one sentence. Both governing visions of the decade — the race and the halt — expect incidents, trust evaluations, know where the compute is, concede that voluntary attestation fails, and never name a responsible person at a developer. The Act is compatible with either future arriving and necessary in both; the strongest attack on this page is to refute any leg of that sentence, quote by quote, in a disposition — each leg above carries its source.
The primaries: the Action Plan (library, read in full) · the AI Futures corpus (the forecasters’ arithmetic). The owners this page leans on: the census · the half-statute page · comparative officer liability · the same conduct · paths to enactment.