Why the disparity exists — twelve explanations, argued rather than assumed
The same conduct, prosecuted establishes the fact: American law reaches individuals for unauthorized computer access with announced exposure from ten years to four hundred and forty, and reaches the officer of a frontier developer not at all.
This file asks why, and refuses to answer it in one sentence. Twelve explanations follow. Several are exculpatory. Several are incompatible with each other. They are drawn deliberately from across the political spectrum, because an explanation a reader already believes is the one that gets them to the evidence — and because a file offering only villains would be answering a different question from the one asked.
Who this is about
About: the officers of frontier developers — companies training above 10²⁶ operations.
Not about: open-source contributors, startups, researchers, ordinary deployers and API customers, or users. On deployers precisely, see the case.
And what this file will not do. Pick a winner. Several of these explanations are true at once, and the useful question for a drafter is not which one but which remedy each implies, because they do not all imply the same one.
How to read the twelve
Each explanation gets four things:
The case for it — stated as its strongest advocate would state it, not as an opponent would. What would confirm it — the evidence that would settle the matter. What it uniquely predicts — the thing this explanation says and the others do not. What remedy it implies — because that is the only part a legislature can act on.
The spectrum labels are approximate and offered without endorsement. They exist so a reader can find the argument they already hold and check it, not to sort anyone.
A · There is no artifact, so the law has nothing to grip
Doctrinal · no political valence
The case for it. Criminal law needs an act, an actor, and a mental state. A person at a keyboard supplies all three cleanly. A deployment decision supplies none: it is made across meetings and reviews, by shifting groups, with no document anyone owns. Prosecutors do not decline these cases out of deference. There is nothing to charge.
What would confirm it. The pattern in why a signature works: twenty-eight years where a false certificate existed, twelve months where twenty-nine people died and no comparable document did.
What it uniquely predicts. That creating the artifact — with nothing else changed — would by itself produce enforcement. This is a testable claim, and Sarbanes-Oxley is the natural experiment: personal certification produced prosecutions that had not existed before.
Remedy implied. SEC. 8. A required signature, and nothing more elaborate.
B · The triggering event has not happened
Historical · no political valence
The case for it. Personal-liability regimes in America are reactive without exception. The Steamboat Acts followed exploding boilers. The 1938 Food, Drug and Cosmetic Act followed a mass poisoning. Mine safety followed disasters. Sarbanes-Oxley followed Enron. No frontier-AI mass-casualty event has occurred in the United States, so the mechanism that produced every previous regime has not been triggered.
What would confirm it. The historical record in house language § 8, which is uniform.
What it uniquely predicts. That the law will arrive after a catastrophe rather than before — and that arguments made beforehand will be ignored regardless of quality.
Remedy implied, and it is uncomfortable for this project. If B is the whole story, the Act arrives when something terrible happens, and the value of drafting now is only that the text exists when the moment comes. That is a smaller claim than this project would like to make, and it may be the true one.
C · Limited liability is working exactly as designed
Market-liberal / conservative
The case for it. The corporate veil is not a loophole. It is deliberate policy, several centuries old, which exists so that capital will fund risky undertakings that individuals would never personally guarantee. The responsible-officer doctrine is the carved exception, and it has only ever been carved industry by industry, slowly, after specific demonstrated need — food and drugs, mines, transport, securities. Frontier AI simply has not been carved yet, and the burden is on whoever wants the carve, as it was every previous time.
What would confirm it. That the doctrine is genuinely narrow and industry-specific, which already a crime, if you are a person shows: the statutes reaching individuals are enumerated, not general.
What it uniquely predicts. That the argument will be won or lost on the specifics of this industry, not on general principle — and that a proponent must show why AI belongs in the carved list.
Remedy implied. A narrow, threshold-limited statute reaching a small covered class — which is what a 10²⁶ threshold produces. This explanation is a friend of the Act’s design, not an enemy.
D · The intent asymmetry is real, and it is not an excuse
Pro-industry · argued in good faith
The case for it. Every defendant in the prosecution gallery did something deliberate. A company whose system does something unanticipated has not. Criminal law is built around chosen wrongdoing, and stretching it to cover unintended outputs of a complex artifact is a serious step that ought to give a legislature pause. The people running these companies are, in the main, trying. Twelve of them published safety frameworks nobody required.
What would confirm it. The frameworks themselves, read in the dossier — several are more demanding than any law requires.
What it uniquely predicts. That a knowledge-based standard would catch nobody, and that any workable duty must run on conduct before release rather than outcomes after it.
Remedy implied — and this is the one the Act adopts. A negligence floor, not a knowledge requirement, attached to the decision to proceed rather than to the harm. The release was intended even where the harm was not. Park was built for exactly this.
E · Prosecutorial economics
Left-institutional, and also a small-government reading
The case for it. An individual computer-crime case is cheap, winnable, and produces a press release. A case against a defendant with unlimited counsel is slow, expensive, and losable. Prosecutors allocate scarce resources rationally. The pattern emerges with nobody intending it.
What would confirm it. The plea data in the gallery: announced maxima enormously in excess of what prosecutors actually sought, which is the signature of a system optimizing for quick resolution.
What it uniquely predicts. That even with a statute, enforcement against well-resourced defendants would be rare — so the deterrent would have to come from the duty’s existence rather than from frequent prosecution.
Remedy implied. Bright-line, documentary offenses that are cheap to prove: a missing signature is provable from an empty file, which is why this Act’s offenses are drafted as record-based rather than harm-based.
F · Causation is genuinely diffuse
Technocratic
The case for it. CFAA harm is traceable to one actor. Harm downstream of a general-purpose model passes through deployers, integrators, professionals and users, each adding decisions. Legal systems handle diffuse causation badly everywhere — this is not special pleading by anyone.
What would confirm it. That the same difficulty appears in unrelated fields: pollution, pharmaceuticals, financial contagion.
What it uniquely predicts. That any harm-based statute will fail on proof, whatever its penalties.
Remedy implied. Seat the offense in conduct rather than consequence — which is the design choice behind SEC. 5(b) and the reason the cross-examination opens on an intrusion rather than a death.
G · Federalism, and the preemption shadow
Constitutional
The case for it. States that might act are drafting under threat of federal preemption; every live federal vehicle contains a preemption clause. Congress moves slowly on anything contested. The result is not that nobody wants to act — it is that the actors who want to are unsure they may.
What would confirm it. The preemption analysis in the drafting record, and the enacted family’s conspicuous avoidance of anything that looks like a design mandate.
What it uniquely predicts. That state statutes will cluster in the safest available space — transparency and reporting — which is precisely what the census found across six regimes.
Remedy implied. Draft into the savings clauses rather than against the preemption clauses. The Act already does this, which is why its offenses are framed as generally applicable criminal law and false-statement offenses.
H · Nobody has put the question to a legislature
Institutional
The case for it. A member reaching for a model reaches for what exists, and what exists is entity-level duties. Counsel drafts; members originate. Staff time follows the bills in front of them. A question absent from the discourse produces no answer regardless of how much expertise is available.
What would confirm it. The commentary sweep: a dedicated academic gap analysis enumerated twenty-six deficiencies in California’s statute and personal accountability was not among them. Three law firms briefed clients on exposure and none mentioned personal exposure, because there is none to mention.
What it uniquely predicts. That the vacancy persists among people acting in complete good faith, and would be closed by the question being asked well once.
Remedy implied. Exactly what this project is: a drafted answer, published where a staffer can find it.
I · The grammar removes the person before anyone gets to the law
Cultural / linguistic
The case for it. House language sets it out: a vocabulary in which systems think, decide and go rogue supplies a culprit who cannot be charged. English has an active construction for the machine and none for the person shipping it. A debate conducted in that grammar cannot arrive at a defendant.
What would confirm it. The frontier statutes’ own drafting: six regimes, and the only human noun in any of them is Connecticut’s officers and directors, appearing as recipients of a report.
What it uniquely predicts. That the vacancy will appear even in documents whose purpose is to find gaps — which is what the gap analysis did.
Remedy implied. No statutory remedy at all. This one is answered by writing differently, which is why the language rule is a repository file and not a preamble.
J · The national-security frame makes personal liability expensive to propose
Hawkish / geopolitical
The case for it. A legislator who proposes criminal exposure for American frontier officers can be told, immediately and publicly, that they are handing the lead to Beijing. That is a real political cost and it is paid up front, whatever the merits.
What would confirm it. The frequency of the argument in hearings and in industry submissions.
What it uniquely predicts. That the objection will be raised in security terms rather than economic ones, and by people not otherwise engaged with the technology.
Remedy implied, and it has an answer. PRC Criminal Law art. 31 already imposes dual punishment — the persons directly in charge and other directly responsible persons are reachable in China in a way they are not in the United States. See the comparative provisions. The competitor being invoked has the rule already.
K · This is the ordinary operation of corporate power, and AI is not special
Left-structural
The case for it. Capital has always externalized liability: through the veil, through subsidiaries, through arbitration, through settlement without admission. Treating frontier AI as an anomaly mistakes a rule for an exception, and a statute aimed at one industry leaves the mechanism intact everywhere else.
What would confirm it. That the same pattern appears in unrelated sectors with no technological novelty at all.
What it uniquely predicts. That closing this gap will produce the next structure — new corporate forms, relocated decision-making, indemnities — rather than compliance.
Remedy implied, and the Act takes it seriously. Anti-evasion drafting: the companion is written against a lineage of liability-evaporation structures, and SEC. 4’s authority test is deliberately indifferent to titles and corporate form for this reason.
L · Lobbying and access
Left · and placed last deliberately
The case for it. The industry is extraordinarily well resourced, employs former officials, funds research and trade associations, and files on the dockets that matter. Who actually files counts the room: twenty-one of fifty-one comments from industry, four from the patient side.
What would confirm it. Disclosure filings, revolving-door records, and the composition data already gathered.
What it uniquely predicts. That the vacancy would persist even after the question was asked well — which is the one prediction that distinguishes L from H, and the one this project is, unintentionally, a live test of.
Remedy implied. Nothing this Act contains. Access reform is a different statute and a different project.
Why it is twelfth rather than first. It is real, and it is the explanation a hostile reader expects, discounts on sight, and uses to file the entire project under grievance. Arriving after eleven explanations that require nobody to be a villain, it lands as one factor among many — which is both more honest and more persuasive than leading with it. A reader who thinks L is the whole story should notice that A, B, C and H each predict the same vacancy with no bad actor anywhere.
The surveys agree — added 23 August
Two independent surveys of the 2026 landscape now state this file’s finding in their own words. CSIS, comparing nine frameworks across two continents (Caroli & Mehta, 3 Aug 2026, ⚠ P): “Many of the frameworks do not include the appointment of senior personnel to oversee the safety framework … This may demonstrate a reluctance by policymakers to impose personal accountability by law, as well as pushback from an industry comfortable with self-implementation of safety requirements but opposed to highly prescriptive governance requirements that delineate specific top-level managers responsible for the safety of models.” And CDT, surveying the year’s most active legislative area (20 Aug 2026, reusable with credit): across the chatbot-safety wave, “[l]iability falls only on the direct deployers … providing protection to original developers of foundation models.” The law reaches down, not up — measured across 146 bills. Explanation H predicted that the vacancy persists among people acting in good faith until the question is asked well once; two surveys have now documented the vacancy without either treating it as a question. The same CSIS report supplies the restraint corollary: on the one documented voluntary withholding of a frontier model, “[n]o authority could have compelled that decision … Only a legally binding regime can guarantee such restraint as a matter of right rather than goodwill.”
The record answers, under oath — added 24 August
Two congressional hearings, read in full, now carry the disparity in sworn testimony — the asymmetry conceded, the immunity requested, and the state-side answer, all at the same tables.
The asymmetry, conceded by a proponent of loosening it. Rush Doshi (Council on Foreign Relations; formerly the NSC’s China directorate), to House Homeland Security on 17 March 2026: “We currently treat American AI companies, like Anthropic, with more regulatory scrutiny than we treat Chinese ones, like DeepSeek.” His remedy runs the other way — restrict the foreign, relieve the domestic — but the sentence concedes the frame this file documents: scrutiny is being assigned by flag, and the argument is over which direction to level it. This Act’s answer is that scrutiny follows the risk and the power over it, whoever holds the flag.
The immunity, requested in terms. Kinsey Fabrizio, president of the Consumer Technology Association, sworn before House Oversight on 17 September 2025: CTA “has urged Congress to adopt a 10-year pause on enforcement of state and local AI laws,” and the federal framework sought is “tech-neutral,” “preemptive,” “risk based” — and “removes liability for companies that are compliant.” The trade association that owns CES asked Congress, under oath, for a design in which compliance purchases immunity. That is explanation L speaking in its own voice, and it is the precise inverse of this Act’s rule, under which compliance is the duty and immunity is nobody’s to purchase.
The state-side answer, same table. Nicol Turner Lee (Brookings), the minority witness at the same hearing: the rejected moratorium “would have threatened states’ rights and the public interest”; since January, “over a hundred measures across 38 states have been enacted to law”; and the sentence that prices the whole dispute — “AGs are trying to figure out ways to keep our grandmothers safe from AI. They are not necessarily trying to compete against China.”
(Sources: House Homeland Security Subcommittee on Cybersecurity and Infrastructure Protection, Serial 119-42, 17 Mar 2026; House Oversight Subcommittee on Cybersecurity, IT, and Government Innovation, Serial 119-49, 17 Sep 2025 — the transcripts, read in full 24 Aug, held in the project library; rows at the verification record § 6.)
What a drafter should take from twelve explanations
They do not all imply the same remedy, and that is the useful finding.
- A, E and F all point to the same drafting choice: documentary, conduct-based offenses, provable from an absence rather than from a causal chain.
- C and G point to narrow scope: a high threshold, a small covered class, drafted into the savings clauses.
- D points to a negligence floor rather than a knowledge requirement.
- H and I are not answered by statute at all, but by asking the question and writing differently.
- B and K are cautions rather than remedies — that the law may arrive only after a catastrophe, and that closing one gap invites the next structure.
- J has a direct answer already on the books elsewhere.
- L is outside this Act’s scope entirely.
Every one of those is what this Act already does, which is either evidence the design is sound or evidence this file was written by the same person who wrote the Act. It was. A reader should weigh it accordingly, and the explanations are set out above in their strongest form precisely so that weighing is possible.
Corrections to the project contact — especially from anyone who holds one of these explanations and thinks it has been stated weakly. They enter the errata register with the fix attached and permanent credit.