Two Greek thinkers compare several law scrolls with one federal scroll beside a mechanical AI head.

One Federal AI Rulebook

California, New York and Illinois regulate the catastrophic risks of the most capable AI models, and federal law says nothing on them ^[2]. A ten-year pause on state enforcement failed in the Senate by 99 votes to 1, and on 23 July 2026 two House members proposed the narrower FRONTIER Act ^[1]. It sets federal duties for the largest developers and displaces the three state statutes, which safety advocates such as Americans for Responsible Innovation contest ^[4].

10 developers or so train models above the computing threshold of the state statutes, by estimate, since no register exists ^[2]
6 of the largest would have to retain a licensed outside examiner with access to their records ^[1]^[3]

The proposal

Replace the state statutes on the catastrophic risks of the most capable AI models with one federal regime, as the FRONTIER Act proposes.

The debate

The sponsors, from both parties, say one national standard with outside audits and incident reports protects the public from catastrophic risk and spares developers a patchwork of state rules. Safety groups welcome the audits but object that the preemption takes effect at once, while the market for licensed examiners takes years to build, and that it never expires. Some critics also fear the clause could reach state rules on patient and consumer safety, which the sponsors deny. The bill was referred to two House committees on 23 July 2026 and has not been taken up.

“This legislation will protect Americans from catastrophic risk, provide developers with clear rules of the road, and ensure the United States remains the global leader in AI.”

argues for: Jay Obernolte, Republican, California Press release, 23 July 2026
“The bipartisan FRONTIER Act delivers commonsense transparency and independent oversight for the largest AI developers while giving them a single, clear national standard to build on.”

argues for: Lori Trahan, Democrat, Massachusetts Press release, 23 July 2026
“There’s also no sunset of the bill’s preemption provision, permanently tying the hands of states on AI laws.”

argues against: Brad Carson, Americans for Responsible Innovation Statement, 23 July 2026

The analysis

One Federal AI Rulebook

36% of the weight is for, 64% against.

An argument's weight combines three questions: how big the effect is, how certain it is and how much the good at stake matters. The larger the slice, the more the argument weighs.

For

Against

One federal AI rulebook is, on the current analysis, probably a poor proposal, and a far narrower one than the debate around it.

Likely worse

Likely worse: Under the assumptions used, the proposal brings more than it costs in 25 of 100 model runs.

Overall effect: small

How strongly the proposal works overall, in five steps from very small to very large, comparable across all debates. The ring is always full, even when the overall effect is small; this step says how large it is.

The analysis leans against the bill, for a reason neither side puts first. The outside verification the sponsors present as its strength is a standing cost that outweighs everything else in the debate. What that verification buys depends on how likely a catastrophic incident is, which nobody knows; at a somewhat higher chance than the analysis assumes, it would pay for itself. The sponsors' other argument, one rulebook instead of three, holds but is small, because the three statutes share a template and bind few companies. The critics' central objection, that states can never tighten the rules, is real and carries no figure, because no state has yet tried to go further.

In detail

All arguments

Arguments for2 arguments

Safeguards checked from outside

17points

State law has the largest developers publish their own safety plans and report incidents. The bill adds a licensed outside examiner with access to unredacted records at any time, and gives the Secretary of Commerce power to halt a model that presents an imminent catastrophic risk. What that is worth depends on how likely such an event is, and nobody knows.

Value 7 · Catastrophic harmImpact 8Plausibility 3

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is harm from the events the bill is written for: a single incident with fifty or more people killed or seriously injured, or damage above one billion dollars, to which one of the most capable models contributes ^[1]. Most of what can be estimated at all is destroyed property and interrupted business, which is money, and part of it is lives, which this site places at the top of the scale. The value therefore lies between the two, nearer to the money, because the one documented comparison is a cyberattack whose toll was counted in dollars. No extra weight is given for the novelty of the technology. The value is above the middle of the scale: mostly money destroyed, partly lives.

The bill defines a catastrophic risk by its floor, fifty casualties or a billion dollars of damage in one incident ^[1]. The costliest cyberattack on record did about ten billion dollars of damage worldwide in 2017 ^[9]. For this evaluation the chance that an event of that class occurs in a given year with one of these models contributing materially is set at two percent, in a range from half a percent to five, and its cost at about 23 billion dollars, in a range from the bill's floor to 120 billion. That is an expected 460 million dollars a year. State law already has developers publish a safety framework and report incidents ^[2]^[8]; what the bill adds is verification every six months by a licensed outside organization for the largest developers, an annual audit, and the power to suspend a model for up to ninety days ^[1]^[3]. The share of the expected harm that those additions avert is set at a tenth, in a range from two to twenty-five percent. The result is about 46 million dollars a year, and between 6 and 120 million once the ranges are taken together short of their joint extremes. The Impact is slightly below the cost of the verification that buys it, and all three of its factors are settings.

▸ Show calculation ▾ Hide calculation

	Chance a year of a catastrophic incident, as the bill defines it, with a frontier model contributing Setting, range 0.5 to 5 percent: no such incident has occurred; the bill's floor is fifty casualties or one billion dollars of damage [1] [1]	2 %	0.02 a year
×	Cost of such an incident Setting, range 0.9 to 100 billion euro-equivalent (about 1 to 120 billion dollars): the costliest cyberattack on record did about ten billion dollars of damage [9], 8.6 billion euro-equivalent at 1.16 dollars to the euro; events with mass casualties would cost more [9]	20 billion euro-equivalent (about 23 billion dollars)	400 million euro-equivalent a year
×	Share averted by outside verification and suspension orders beyond what state law achieves Setting, range 2 to 25 percent: state law already requires published frameworks and incident reports [2]; examiners are chosen by the developer [4] [2][4]	10 %	40 million euro-equivalent a year
÷	Normalised Impact scale of this evaluation	5 million euro-equivalent a point	8

Score 8 Impact × 7 Value × 3 Plausibility ÷ 10 = 17 of 100

Nothing here has been measured, because the event has not occurred and the verification regime does not exist. The comparison is state law, under which developers write and publish their own frameworks and nobody outside checks them ^[2]^[8]. The chain has three links: outside examiners with full access find weaknesses the developer missed or discounted, the developer or the Secretary acts on them, and an incident that would have happened does not. The first link is supported by experience with financial audits and safety inspection in other industries and has no evidence for this one. Against it stands that developers choose their own examiner, which critics see as a conflict of interest, and that the decision to release a model stays with the developer ^[4]. The examiner market also has to be built first, which takes years ^[4]. Nothing suggests the link runs the other way. The Plausibility is low: the mechanism is ordinary, the probability and the averted share are set rather than known, and the examiners do not yet exist.

Evidence: Mechanism · plausibility at most 6 Evidence of effect: Mechanistic · at most 6 Placement: Chain open · plausibility 3–3.5

Open: The semiannual reports to the Under Secretary and the incident reports filed in the first two years would show how often outside examiners find what developers had not disclosed, which is the first link of the chain.

One rulebook instead of three

9.5points

California, New York and Illinois each have a statute on the catastrophic risks of the most capable models, written from the same template with different deadlines, authorities and penalties. The bill replaces the three with one. The saving is real and small, because only about ten companies are in scope.

Value 5 · Enforcement costImpact 2Plausibility 9.5

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is money spent on meeting the same duty three times over rather than once. It is priced like any other money a company spends on running a rule, and it is a genuine cost rather than a transfer: the legal hours, the parallel filings and the separate incident procedures are consumed. Whether the developer or its customers carry it makes no difference to the weight. What the duties themselves achieve is not counted here, because the federal bill keeps them and adds to them. The value is the middle of the scale, the level this site uses for money spent on running a rule.

The bill displaces state obligations on developers in three subjects: transparency about catastrophic risk, outside audits and verification, and incident reporting ^[1]. Its sponsors name the statutes it is aimed at, California's SB 53, New York's RAISE Act and Illinois's SB 315 ^[1]. All three apply from the same computing threshold and differ in the details: California allows fifteen days to report an incident and the other two allow 72 hours, Illinois requires an annual audit from 2028 and California none, and the penalties differ ^[2]^[8]. About ten developers train models above the threshold, in a range from six to sixteen; no register exists yet, so the count is an estimate. The duplicated work is put at about 1.2 million dollars each a year, in a range from 350,000 to 3.5 million: legal review of three statutes, three sets of filings, and incident procedures built to the strictest deadline. That gives about 12 million dollars a year, and between about 5 and 35 million once both ranges are taken together short of their joint extremes. Every rule on companies that merely use these models stays with the states, so nothing is saved for the hospitals, insurers and employers that the wider debate is about ^[1]. The Impact is the smallest in this debate: the patchwork the bill removes consists of three similar statutes binding about ten companies.

▸ Show calculation ▾ Hide calculation

	Developers training models above the computing threshold of the three state statutes Setting, range 6 to 16: no register exists; the federal one would be the first [2]	same threshold in California, New York and Illinois	10 developers
×	Duplicated compliance cost each a year Setting, range 0.3 to 3 million euro-equivalent (about 0.35 to 3.5 million dollars): legal review of three statutes, three sets of filings, incident procedures built to the strictest deadline; low because the statutes share a template [2][8]	1 million euro-equivalent (about 1.2 million dollars at 1.16 dollars to the euro)	10 million euro-equivalent
×	Weight of a dollar in company budgets the standard weight this site uses for business money	1.0	10 million euro-equivalent
÷	Normalised Impact scale of this evaluation	5 million euro-equivalent a point	2

Score 2 Impact × 5 Value × 9.5 Plausibility ÷ 10 = 9.5 of 100

That one statute instead of three saves the duplicated work follows from the text: a state may neither adopt nor enforce an obligation on developers in the three subjects once the bill is law ^[1]^[2]. The comparison is the three state statutes as they stand. What is uncertain is the size, and neither part of it has a source: nobody publishes how many developers are above the threshold, and nobody publishes what parallel compliance costs them. Because the three statutes share a template, a developer that meets the strictest may already satisfy most of the rest, which is why the figure per company is low. These doubts sit in the range around the figure and are not counted a second time here. The remaining doubt is the reach of the clause: it speaks of new substantive obligations, and critics expect its edges to be argued in court ^[4]. The Plausibility is very high: the saving follows from the displacement itself, and its size is carried in the range.

Evidence: Mechanism · plausibility at most 9.5 Evidence of effect: Definitional

Arguments against2 arguments

Verification costs more than state law

48points

The bill keeps every duty the three state statutes impose and adds one they do not have: the largest developers must retain a licensed outside organization that examines their safeguards continuously and reports twice a year. That is a standing cost, and it falls on about six companies.

Value 5 · Company budgetsImpact 10Plausibility 9.5

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is money that developers spend on examiners, internal staff to serve them and the annual compliance audit. It is priced like any other money in a company budget, at the middle of the scale. It is a real cost rather than a transfer, because the examiners' hours are consumed. Whether the developers, their investors or their customers carry it makes no difference to the weight, and no discount is applied because the companies concerned are very large. What the examination achieves is counted on the other side. The value is the middle of the scale, the level this site uses for money in company budgets.

A developer with more than five billion dollars of revenue and at least ten billion dollars of spending on model development over three years must retain a licensed verification organization, give it unredacted records at any time and receive an assessment every six months ^[1]^[3]. About six companies meet both thresholds, in a range from four to eight. No such examiner exists yet and none has quoted a price. The cost is put at about 9.3 million dollars a year for each developer, in a range from 3.5 to 23 million: a standing team of specialists with security clearance for model weights, plus the developer's own staff to serve it. That is about 56 million dollars a year. The annual compliance audit for the next tier adds little, because Illinois requires an audit from 2028 anyway; one additional year for about ten developers at about 1.2 million dollars each comes to 2.3 million a year over the five years ^[2]^[8]. Together that is about 58 million dollars a year, and between 23 and 140 million once the ranges are taken together short of their joint extremes. The Impact is the largest in this debate, five times the duplicated paperwork the bill removes.

▸ Show calculation ▾ Hide calculation

	Developers above both thresholds of the top tier Setting, range 4 to 8: the thresholds are in the bill [1][3], the count is not published [1][3]	more than 5 billion dollars revenue and at least 10 billion dollars development spending in 36 months	6 developers
×	Cost of continuous outside verification each a year Setting, range 3 to 20 million euro-equivalent (about 3.5 to 23 million dollars): a standing specialist team with access to unredacted records, plus internal staff; no examiner exists yet and none has quoted a price [4] [4]	8 million euro-equivalent (about 9.3 million dollars at 1.16 dollars to the euro)	48 million euro-equivalent a year
+	Annual compliance audit in the one year before Illinois requires it anyway Illinois requires an annual audit from 2028 [2]; only the earlier start is additional [2]	10 developers × 1 million euro-equivalent (about 1.2 million dollars) ÷ 5 years	50 million euro-equivalent a year
×	Weight of a dollar in company budgets the standard weight this site uses for business money	1.0	50 million euro-equivalent a year
÷	Normalised Impact scale of this evaluation	5 million euro-equivalent a point	10

Score 10 Impact × 5 Value × 9.5 Plausibility ÷ 10 = 48 of 100

That the cost arises follows from the text: the duty to retain an examiner is unconditional for developers above the thresholds, and a violation carries a civil penalty of up to one million dollars a day ^[1]. The comparison is state law, which requires no continuous outside verification; Illinois alone requires an annual audit, from 2028 ^[2]. What is uncertain is the price, since the market for these examiners has to be created first ^[4]. That doubt sits in the range around the figure and is not counted a second time here. The Under Secretary may raise the thresholds by rule and may not lower them, so the number of companies in scope can only fall ^[1]^[2]. The remaining doubt is timing: licensing rules are due within 180 days of enactment, and examiners will take longer to appear ^[1]^[4]. The Plausibility is very high: the duty is written into the bill, and only its price is open.

Evidence: Mechanism · plausibility at most 9.5 Evidence of effect: Definitional

States cannot tighten the rules

0points

The bill has no expiry date. On its three subjects no state may add a duty for developers, and the federal official in charge may raise the thresholds that decide who is covered and may not lower them. If the federal rules prove too weak, the only remedy is another act of Congress.

Value 6 · Federal latitudeImpact 0Plausibility 4

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream would be the capacity to respond when a rule proves too weak: three legislatures that have each acted within a year, against a Congress that has enacted no statute on the subject. This site places that in the class it uses for the working order of public institutions. What would be priced is the capacity to respond rather than any particular rule it might produce, and nothing here treats state authority as valuable in its own right. The earlier discussion draft limited its preemption to three years; the bill as introduced does not ^[2]^[5]. The value would sit in the middle-upper part of the scale, at the level this site uses for the working order of public institutions.

The bill bars states from adopting or enforcing developer obligations on catastrophic-risk transparency, outside audits and incident reporting, without a sunset ^[1]^[2]. Everything else stays with the states: rules on companies and agencies that use these systems, laws protecting minors, laws that apply to all businesses, and procurement ^[1]. The chatbot safety laws of 2026 and the rights over automated decisions in hiring, lending and insurance are therefore untouched. On the three subjects the federal side is not empty either, since an Under Secretary writes minimum requirements by rule within 180 days ^[1]. What remains is narrower: a state may not go beyond the federal rule, and coverage can only shrink, because thresholds may be raised and not lowered ^[2]. No quantity can be derived for that, since no state has yet tried to go further than the federal bill does. The argument is therefore included and enters no total. The Impact is not quantified: the loss is a response that no state has yet attempted.

That states lose the power follows from the text, and it takes effect on enactment ^[1]^[4]. The comparison is the present position, in which California, New York and Illinois legislated on these subjects within about a year of each other ^[2]^[8]. Whether the loss costs anything depends on a future in which the federal rules prove inadequate and a state would have acted where Congress and the Under Secretary did not. Critics add that the wording could reach further than its sponsors intend, to state rules on safety for patients and consumers, which is a claim about how courts will read the clause and not something the text says ^[4]. The sponsors' own summary states the opposite ^[1]. Nothing suggests the link runs the other way. The Plausibility would be below the middle: the loss of power is certain, and whether it will ever matter is open.

Recorded, not yet scored: The bill removes, without a sunset, the states' power to add developer obligations on catastrophic-risk transparency, auditing and incident reporting. No state has yet tried to go beyond what the federal bill requires, so there is no episode from which a quantity could be derived.

Evidence: Mechanism · plausibility at most 6 Evidence of effect: Mechanistic · at most 6

How robust the result is

Under the assumptions used, the proposal brings more than it costs in 25 of 100 model runs.

gains more than it costscosts more than it gains

Each dot stands for 200 of the 20,000 model runs.

This evaluation scores the FRONTIER Act as introduced in the House in July 2026 and compares five years under it with five years of the three state statutes. The result turns on one number that cannot be sourced: the yearly chance of a catastrophic incident with a frontier model contributing, set at two percent. How much of that harm outside verification averts beyond what state law already achieves is a setting too, a tenth. The value of being able to tighten the rules later is recorded without a figure. At two percent the bill lies ahead in roughly one of four runs.

Show the distribution of the model runs

20,000 simulated futures

Likely worse · 25 %

In 25 of 100 simulated futures the proposal brings more than it costs.

No effect is known exactly; every argument has a range. The calculation draws a possible size for each argument from its range 20,000 times and counts how often the benefits exceed the costs.

Net effect in points per future. Left of zero the drawbacks outweigh the benefits, right of zero the benefits win. Eight in ten futures fall between −70 and +18 points.

Scale and assessment

Scale of this evaluation: Normalised Impact, unitless and calibrated to this topic. For comparison, one point here corresponds to about 5 million euro-equivalent a year.

Size class:very small

Effect small (−0.11) · 36 % of the effect is positive

Areas show individual argument weights. The positive share of the effect nets out money transfers between the two sides; the probability of a positive net effect is determined separately.

How we score →

Review and sources

Last reviewed by Claude Opus 5.5 · October 29, 2026 · 9× AI, 2× human

October 29, 2026AI reviewClaude Opus 5.5record updated
Page texts for the new page layout (draft G) with quotes from primary sources; numbers unchanged.
October 21, 2026AI reviewClaude Fable 5.1record updated
Money amounts in argument texts stated in dollars per style-en.md; scores, chains and balance unchanged
October 20, 2026AI reviewClaude Fable 5.1record updated
Summary and scale note rewritten to the style of 20 Oct 2026; numbers unchanged
October 14, 2026AI reviewClaude Opus 5.5record updated
Prüfbericht 11.10. gegen v4: Maßnahmentext nennt kein absolutes Fehlen eines AI-Bundesgesetzes mehr (TAKE IT DOWN Act, P.L. 119-12); der con-1-Scopewiderspruch ist mit v4 entfallen.
October 12, 2026Assessed by a human, implemented by AIClaude Opus 5.5re-scored
v4 nach Julians Entscheidung vom 12.10. (bewertet wird der aktuellste Entwurf): umgestellt auf den FRONTIER Act, H.R. 9925, eingebracht Juli 2026 (nur Entwicklerpflichten zu Katastrophenrisiko-Transparenz, Audit, Vorfallmeldung; kein Ablaufdatum; Anwenderrecht der Staaten bleibt). con-1, con-3, pro-2, pro-3 entfallen; pro-1 auf drei Landesgesetze geschrumpft (I 0,6 → 0,01); neu pro-4 (externe Prüfung, Setzungen) und con-4 (Prüfkosten); con-2 als bewertung_offen. r 0,70 → 0,36, Bilanz P(D>0) 0,214 = schlechter; k 5 → 200.
October 12, 2026AI reviewClaude Opus 5.5record updated
Prüfbericht 12.10. in place umgesetzt: i_spanne an pro-2 (0,06–1,0) und pro-3 (0,2–2,4), con-2 auf 0–0,26; con-1 Wert auf grundrechte mit neuer V-Begründung; con-2 Rahmenpapier richtig wiedergegeben (keine neue Regelsetzungsbehörde); Zahl 29 mit Quelle [8] belegt; Revisionserzählung, Mechanik-Sätze und Schreibweise bereinigt; Rechenweg con-3; keine Zahl geändert, kein bilanz-Block (Probelauf 0,516 = ausgeglichen gegen r 0,70 = besser).
October 2, 2026Assessed by a human, implemented by AIClaude Opus 5.5re-scored
Regel ‚Eintritt und Höhe getrennt' (Julian 02.10.) angewendet: pro-1 P 5 → 9,5 mit neuer i_spanne 0,2–2,0; con-1 von definitorisch auf mechanistisch korrigiert (der Schaden hängt an der Nutzung der Rechte), P 4 bleibt; r 0,63 → 0,70, Kategorie besser (ohne Bilanz-Block, pro-2/pro-3 ohne Spanne).
September 14, 2026AI reviewClaude Fable 5.1re-scored
v2 nach Gesamtprüfung 08.09.: Framework-Primärtext gelesen — Kinderschutzgesetze nicht präemptiert, Vollzug über bestehende Behörden; con-1 1,2 → 0,54 (ohne Companion-Anteil), con-2 0,5 → 0,13 (eine Episode statt Dauerstrom), con-3 auf 0 (Fehllesung + Doppelzählung mit con-1). r 0,37 → 0,63.
September 6, 2026AI reviewClaude Opus 5First evaluation
Created for the English side: scores the statutory framework, not the executive order and litigation strategy now in force.

Evaluations are produced with AI support and reviewed on a schedule for new developments; human passes are marked separately.How we review →

Show all 9 sourcesHide sources

Reps. Lori Trahan and Jay Obernolte: FRONTIER Act: Section-by-Section Summary (21 July 2026). trahan.house.gov
US Frontier AI Legislation Tracker: H.R. 9925, FRONTIER Act: status, preemption clause and comparison with California SB 53, New York RAISE and Illinois SB 315. usfrontierlaw.com
Foundation for American Innovation: The FRONTIER Act Is Congress's Best AI Bill Yet. thefai.org
Americans for Responsible Innovation: FRONTIER Act Improves on Prior Proposal, but Requires Critical Changes. ari.us
Roll Call: Bipartisan AI draft proposes three-year preemption of state laws. rollcall.com
The White House: National Policy Framework for Artificial Intelligence — Legislative Recommendations (March 2026). whitehouse.gov
Tech Policy Press: Where State AI Legislation Stands Half Way Into 2026. techpolicy.press
Future of Privacy Forum: Frontier AI Goes Federal: How the Great American AI Act Compares to State Laws. fpf.org
WIRED: The Untold Story of NotPetya, the Most Devastating Cyberattack in History. wired.com