Americans for Responsible Innovation (principal author Iskandar Haykel) · 2026
Americans for Responsible Innovation — Responsible Innovation at the Frontier: A Blueprint for Federal AI Governance
ARI Blueprint
Published August 6, 2026 by Americans for Responsible Innovation, the advocacy group led by former congressman Brad Carson, the blueprint is a position paper rather than legislative text; ARI says statutory language will decide its position on any bill. It starts from a gap in the design the state frontier laws and the developers' own federal blueprints share: developers publish safety frameworks, but they write those frameworks themselves, can revise them at will, and nothing guarantees they are adequate. ARI's answer is three functions it calls inseparable. Standards: each covered developer publishes a framework addressing five statutory risk domains (chemical, biological, radiological, nuclear, and explosive threats; offensive cyber; automated AI R&D; harmful manipulation, including foreign influence operations; and autonomy, including misalignment and loss of control), and the regulator sets minimum adequacy standards on a fixed cycle that can rise but never fall below the previous cycle's floor. Assurance: the government examines every covered developer from the outset, both for compliance with its framework and for whether the framework is adequate, modeled on the Federal Reserve's standing supervision of the largest banks; accredited private verifiers may later share the compliance work, domain by domain and only once the regulator certifies enough capacity exists, but the adequacy judgment is never delegated. Transparency: confidential quarterly filings on the automation of AI research and development, with a material-change filing within four business days, adapted from the SEC's 10-Q and 8-K regime. Behind all three sits a residual emergency authority: named senior officials may halt development, internal deployment, or release, order mitigations, or recall a model, with orders lapsing after 72 hours unless the government asks a court to extend them. Coverage is deliberately narrow, reaching only developers that have trained a model at 10^26 operations and spent $100 million on training in the past year, and it includes internal use. Federal compliance satisfies a state requirement only where it covers the same conduct and is at least as protective, states may legislate above the floor, compliance confers no safe harbor, and the whole regime sunsets after nine years unless Congress reauthorizes it.
Key Provisions
Coverage: developers that have trained, or begun training, a model at 10^26 operations and have spent at least $100M on AI training in the past year; the regulator reviews the compute threshold on a fixed cycle and, with CAISI, should move to a capabilities-based definition; obligations reach internal as well as external deployment
Published safety frameworks addressing five statutory risk domains (CBRNE, offensive cyber, automated AI R&D, harmful manipulation including foreign influence operations, and autonomy including loss of control), each stating how risk is evaluated and why that is sufficient, the thresholds that trigger mitigation, and when federal authorities are notified; a named accountable officer; frameworks filed with the regulator and published in a public catalog
System cards with evaluation results published alongside public releases, and risk reports published on a regular basis after deployment
Federal minimum adequacy standards reviewed on a fixed cycle, informed by developer practice, CAISI analysis, independent research, and examination findings, but set by the regulator alone; standards may never fall below the previous cycle's, and risk domains can be added but never removed
Residual emergency authority, held only by named senior officials, to halt development, internal deployment, or release, order mitigations, or recall a model; orders take effect immediately, lapse after 72 hours unless the government petitions a court, and survive only on substantial evidence of present or imminent severe harm; frameworks must themselves state the conditions for stopping, and a disputed mitigation means a halt until the dispute is resolved
Government assurance from day one: federal evaluation teams test models against the standards and each framework's thresholds, while a standing oversight team examines safety practices; incident reports, statutorily protected whistleblowers, and examiner findings can each trigger a for-cause examination
Accredited independent verification organizations (IVOs) phased in only where the regulator certifies capacity: first government alone, then government plus a developer-chosen IVO, then a developer-chosen and a government-assigned IVO, with the government keeping a share of examinations in every domain and supervising IVOs on the PCAOB model
Verifier independence: examination fees routed through a regulator-administered fund, a cap on any IVO's revenue from one developer, mandatory rotation of IVOs and government teams, revolving-door limits, and review of divergent assessments; IVOs lose liability protection once their examinations are load-bearing
AI R&D automation disclosure: confidential quarterly filings and four-day material-change filings anchored on auditable measures such as the share of compute spent on autonomous work, with a statutory test of meaningful human oversight and review-sampling rates; public summaries only for the safety, evaluation, and oversight category; penalties for knowingly false statements or lost records, protection for good-faith corrections
Preemption by compliance equivalence only: federal compliance satisfies a state requirement that covers the same conduct and is no more protective; states remain free to legislate above the floor, and compliance creates no immunity, safe harbor, or presumption against liability
Nine-year sunset unless reauthorized, with staggered efficacy reviews of each function and a GAO review before expiry
Regulatory Philosophy
Supervision, not self-certification. ARI treats frontier labs the way bank regulators treat the largest banks: the government sets the standard, examines continuously, and never hands the final judgment to anyone else. Where the FRONTIER Act and Anthropic's text build oversight around private evaluators the developers retain, ARI makes the government the primary examiner from the start and treats a private verification market as something that may never emerge, writing the plan so that it works either way. Its central device is the ratchet: standards set by the regulator, informed but not authored by industry, that can only move up. It offers this as a bargain — tight obligations on a handful of developers, a federal floor rather than a ceiling, and a sunset — and presents clear, rules-based authority as the alternative to what it describes as the administration's current practice of preventing frontier releases on unspecified criteria.
Developers that have trained a model at 10^26 operations and spent at least $100M on AI training in the past year
Why it lands there
Falls entirely on the handful of frontier developers and reaches into their internal research pipelines through the R&D disclosure regime. Developers pay examination fees once private verifiers arrive, routed through a regulator-administered fund; the larger cost, a permanent federal examination workforce, falls on the government, and the blueprint does not say how it would be paid for.
ARI's blueprint and the FRONTIER Act start in the same place, published frameworks at the 10^26 frontier with internal use covered, and both add someone outside the developer who judges whether its mitigation is adequate, plus an executive power to stop a model in an emergency. They split on who that someone is. The FRONTIER Act hands ongoing assessment of its largest developers to federally licensed private verifiers from the start and shields them from suit over the models they assess. ARI keeps all examination in government hands until the regulator certifies that a private market can carry it, phases verifiers in domain by domain while the government keeps a share of the work and sole authority over adequacy, and warns that 'there could be nothing worse for frontier AI safety than a robust IVO market that enjoys immunity from its own negligent conduct.' The emergency powers differ in duration: FRONTIER's Commerce Secretary issues 45-day provisional and 90-day renewable orders, while ARI's lapse after 72 hours unless a court extends them. And on the states they point in opposite directions: FRONTIER permanently preempts state transparency, auditing, and incident-reporting rules for frontier developers, while ARI lets federal compliance satisfy only state requirements that are no stricter, leaves states free to go further, and sunsets itself after nine years.
+Makes the adequacy of each developer's framework a government judgment that is never delegated, closing the gap the state laws leave open: a developer can comply perfectly with a framework too weak to matter
+The ratchet, under which standards cannot fall below the last cycle's and statutory risk domains cannot be removed, ends developers' ability to quietly revise their own safety commitments downward
+Does not depend on a private verifier market existing, and when one does, guards it against capture with pooled fees, revenue caps, rotation, revolving-door limits, and no immunity for verifiers whose work is load-bearing, the reverse of the FRONTIER Act's IVO liability shield
+The most detailed treatment of AI R&D automation on the map: auditable measures, a test of meaningful human oversight, and a four-day clock for material changes, aimed at the capability most likely to outrun any fixed review cycle
+Bounds the emergency power tightly: named officials, a 72-hour default lapse, and a substantial-evidence standard, with the conditions for stopping written into each developer's own framework in advance
+The most state-protective federal frontier design: a floor rather than a ceiling, no safe harbor from state liability, and a nine-year sunset
Weaknesses
From the perspective of political opposition
−A blueprint, not a bill: it leaves open which agency holds which function, the size of penalties, and any deadline for reporting incidents, and ARI reserves its position on any actual legislation
−Requires the government to examine every covered developer across five risk domains from day one, a workforce that does not exist today, without saying how it would be hired or funded
−Emergency orders take effect immediately on an executive official's finding and can stay in force through expedited litigation, the kind of executive power over model releases that critics of the current administration's interventions object to
−Requiring both 10^26 operations and $100M in training spend leaves out efficient or smaller developers whose models approach frontier capability until the regulator revises the threshold
−The 'harmful manipulation' domain, which includes influence operations, is less well defined than the others and invites First Amendment challenges
−Compliance-equivalence preemption gives developers little relief from a state patchwork, since any state can still legislate above the floor, removing the main incentive the industry has to support federal law