This memorandum accompanies two drafts: a Minnesota bill written to the 2027 session, and a federal bill of the same structure. It is written by the drafts' author, who is not a lawyer. It states the argument the drafts rest on and, where the argument is contestable, says so.
Read this alongside OPEN-ISSUES.md, which records what is unresolved.
1. Start with Minnesota
The Minnesota draft is the easier bill, and not only because a state session is easier to enter than a Congress. The hardest questions about the federal draft are questions about federal power over states. None of them arise in Minnesota.
| Question | Federal draft | Minnesota draft |
|---|---|---|
| May the legislature direct state agencies this way? | Contested — see §3 | Yes. Ordinary legislative power over state agencies and political subdivisions. |
| Murphy v. NCAA anti-commandeering | Live issue, §105(b) | Does not arise. A state legislature directing its own state is not commandeering. |
| City of Boerne limits on §5 enforcement power | Live issue | Does not arise. Minnesota is not enforcing the Fourteenth Amendment against itself. |
| Personal liability of state officers | Aggressive; draws opposition | Not attempted. Article 1, §19, subd. 9 expressly creates no personal liability. |
| Preemption architecture | The core legal question | Only local preemption, §23 — routine. |
A Minnesota committee is therefore being asked a policy question — should the state bind its own hands as to capability restrictions — rather than a constitutional one. That is the pitch.
The one Minnesota-specific structural question is the Board of Regents, addressed at §6 below.
2. What the drafts do, in one page
They withhold from government a particular power: the power to restrict an AI system or a computational resource because of how capable it is. Everything government can do about conduct it keeps.
The mechanism is four-part.
- A statutory right to acquire, possess, develop, share, access and operate computational resources and AI systems for lawful purposes. Federal §§102–104; Minnesota §§5–7.
- A categorical bar on capability-based restrictions. Not a strict-scrutiny test — a withholding of authority. Federal §105; Minnesota §8. No interest and no tailoring validates such a restriction.
- A weaker, RFRA-modeled test for burdens that are not capability-based: compelling interest, least restrictive reasonably available means, particularized findings. Federal §108; Minnesota §11 (which adds clear-and-convincing proof).
- Anti-circumvention, because the easy route is indirect: pressure on cloud providers, procurement conditions, registration requirements, capability-keyed taxes. Federal §§105(c), 106; Minnesota §§8 subd. 4, 9.
Why categorical rather than balanced. A balancing test concedes the question. Under balancing the issue becomes whether the asserted worry is weighty enough, and every capability restriction proposed so far has been justified by a weighty-sounding worry. An earlier draft used a compelling-interest test for everything; it was replaced because a protection that yields whenever the worry is dramatic is not a protection. This is the drafts' central choice and a committee should be asked to accept or reject it deliberately.
3. Federal authority: the argument, and its weak point
The claim. Congress may enact the federal draft under the Commerce Clause and the Necessary and Proper Clause. Finding (10) sets out the interstate predicate: compute, model weights, inference services and the hardware they run on are bought and sold across state lines; systems are trained in one state, hosted in another and used in every state; providers cannot partition service by state at reasonable cost, so a single state's capability restriction operates on the national market.
Why the drafts do not use §5 of the Fourteenth Amendment. City of Boerne v. Flores, 521 U.S. 507 (1997), held that Congress could not impose RFRA's compelling-interest standard on states under §5. Federal §108 applies a substantively similar standard to states through §101(e)(1). If that rested on §5 it would fail. It rests on the commerce power instead. Counsel must make that argument rather than assume it, because the resemblance to RFRA is the first thing a state attorney general will point at.
The Murphy problem, and what was done about it. Murphy v. NCAA, 584 U.S. 453 (2018), struck PASPA because it told state legislatures what they could not enact. An earlier draft of §105(b) read "No State or political subdivision may enact or enforce," which is close to that phrasing. The section now reads differently: the restriction "has no force or effect," a covered person is "entitled … to acquire, possess, develop, share, operate, and access such a system or resource free of any such restriction," and the subsection says expressly that it regulates the interstate market and confers a right on private persons rather than directing a legislature. That is the structure Murphy said was permissible — federal law regulating private actors, with which conflicting state law is preempted.
This rewrite is the author's drafting and has not been reviewed by counsel. It is the single highest-priority item for a drafting office. If counsel prefers a different formulation, the rest of the Act does not depend on this one.
What is still exposed. Federal §114(d) allows individual-capacity damages against state and local officers with no qualified immunity (§114(k)). Congress imposing personal liability on state officers under the commerce power, with no §5 hook after Boerne, is the most aggressive thing the federal draft does. It is severable under §301(c). Governors, attorneys general and local-government associations will oppose it. An office that wants the Act without that fight can strike §114(d) and §114(k) and lose nothing else. The Minnesota draft does not attempt this at all.
4. The boundary: when is a rule "capability-based"?
This is the question a committee will actually press, because the prohibition is categorical and therefore the definition does all the work. The line the drafts draw is:
A rule that picks out its subject by what a system can do is prohibited. A rule that picks out its subject by what a person does, or by a separately regulated activity or facility, is not.
Definitions: federal §3(4) (capability-based restriction) and §3(17) (technical proxy, closing with a purpose-or-effect catch-all); Minnesota §3, subds. 6 and 15. Savings: federal §§105(e), 105(h), 113(a)–(d), 115(b); Minnesota §8, subds. 6–7, §22.
Five scenarios, worked through. These are the tests named in OPEN-ISSUES.md.
4.1 Mandatory reporting triggered by a demonstrated ability to synthesize a named biological agent
Prohibited as drafted. The trigger is a capability ("scientific or technical ability"), and §105(c) names "a reporting or registration requirement" among the prohibited proxies. Preventing CBRN harm is a recognized compelling interest (federal §109(b)), but §109(i) forecloses using that interest to justify a capability restriction, and §109(a) says the interest "may be furthered by the prohibition, investigation, and punishment of independently unlawful conduct."
What remains available. Everything aimed at conduct: prohibiting the synthesis itself, controlling precursors and synthesis equipment, regulating the laboratories and gene-synthesis providers that perform the work, the screening obligations of those providers, and prosecution of anyone who attempts it. Under federal §113(a) those survive intact.
Whether a committee should accept this. This is the sharpest case, and the drafts answer it the same way they answer everything else: regulate the act, not the tool. An office that finds that unacceptable for biological synthesis specifically should amend §109 rather than dilute §105, and should do so knowingly. The draft's position is that a reporting duty keyed to model capability does not stop a determined actor and does reach every legitimate researcher.
Two savings provisions added after outside review bear directly on this section. Federal §113(c) and Minnesota §22, subd. 3 state that a law reaching a product designed, marketed, or primarily used to produce an independently unlawful result — a nonconsensual intimate image, child sexual abuse material, a forged identification document — regulates that conduct and is not a capability-based restriction. Federal §113(d) and Minnesota §22, subd. 4 state that safety regulation of a vehicle, aircraft, vessel, unmanned aircraft system, medical device, or other separately regulated product is not a capability-based restriction merely because it classifies by degree of automation or by demonstrated performance. Neither permits a restriction on a general-purpose system by reason of its capability.
4.2 Safety testing triggered by autonomous control of industrial machinery
Permitted. The trigger is a deployment — the system is controlling machinery — not a capability. This is the same theory as §116 (federal) and §15 (Minnesota), which impose a written risk-management policy and a human-override guarantee on anyone deploying AI to control critical infrastructure, and which §116(d) expressly declares is not a capability-based restriction.
Line to hold. "Systems that control industrial machinery must pass safety testing" is a rule about an activity. "Systems above N FLOPs must pass safety testing before anyone may run them" is a capability ceiling. The first is fine; the second is void.
4.3 AI functionality inside a regulated medical device
Permitted, and the text says so explicitly. FDA device regulation reaches the device — its safety and efficacy for an intended use — not the capability of software as such. The risk was federal §117(d), which barred conditioning access to an AI system on "professional licensure" and would literally have reached prescription-only AI devices. §117(d) now carries a savings clause preserving requirements that condition "the practice of a licensed profession, or the use of a device or product regulated as to safety or efficacy, on professional licensure or professional supervision," where the requirement regulates the practice or the product rather than capability.
Federal §113(a) also now names "professional conduct" and "professional licensure," and §115(b) preserves equivalent state law.
4.4 Licensing conditions for AI controlling nuclear infrastructure
Permitted. Nuclear facility licensing is the paradigm case of regulating a separately regulated facility. Federal §105(e) and Minnesota §8, subd. 7 preserve laws governing a facility "when that law regulates the facility rather than computation itself," and Minnesota's version adds "or a law regulating a separately regulated facility or activity."
A condition in an NRC license about what software may control a reactor is a condition on operating a reactor. It would only become a capability restriction if it were written to reach the licensee's unrelated computing — for instance, barring a utility from running an advanced model anywhere in its business.
4.5 Cybersecurity rules triggered by autonomous exploitation capability
Prohibited, deliberately, and this is the drafts' strongest case. The trigger is squarely a capability ("cybersecurity ability"), and federal §107 and Minnesota §10 protect defensive use specifically. §107(g) states the reason: "That a capability usable to secure a system is also usable to attack one is not, by itself, sufficient."
The same tooling that finds a vulnerability for an attacker finds it for the defender. A rule keyed to exploitation capability disarms defenders — who comply — and not attackers, who do not. Unauthorized access remains prohibited under §109(c) and every existing computer-crime statute.
Note. §107(b) and Minnesota §10, subd. 2 previously carried an "unless" clause allowing such a restriction if it survived §§105 and 108. That was a null set, since a restriction premised on capability is capability-based by definition. The clause is deleted and the rule is flat.
What the five have in common
The permitted cases (4.2, 4.3, 4.4) all regulate an activity, a product in use, or a facility. The prohibited cases (4.1, 4.5) regulate the tool, in anyone's hands, before anything happens. That is the line. A committee that wants to move any specific case across it should amend §109 and say so on the record, rather than soften the definitions — softening the definitions reopens everything.
5. What the drafts cost
Rights provisions: no appropriation, but not free. Administrative cost for the annual inventories (federal §120(f), Minnesota §21, subd. 6 — both assigned to the attorney general), the capability disclosure (federal §117(e), OMB), and the rulemaking statements (§111). Litigation exposure: a prevailing plaintiff is awarded fees, and the federal bill allows statutory and compensatory damages in defined cases. Statutory damages are now capped at $50,000 per plaintiff per governmental action.
Private duties, three of them. A written risk policy and human override for AI controlling critical infrastructure, with an exemption for businesses under 50 employees unless they serve more than 10,000 customers. Data portability on user request. Notice to a user cut off because of government pressure — on the intermediary federally, on the government entity in Minnesota.
The public-access program is the only direct spending, and it is the item to scrutinise. Its capability floor is measured against a reference set that includes commercial, academic and nonprofit systems, not government systems alone (federal §205(b); Minnesota article 2, §5, subd. 1). That is a materially larger commitment than "parity with what the state runs," and it should be priced against the broader benchmark. Both drafts make the program severable precisely so a committee can remove it without touching the rights. Appropriation figures are left blank.
6. The Minnesota Board of Regents question
Article 2 asks the Board of Regents, through the Minnesota Supercomputing Institute, to run the public-access program. It says "is requested to" throughout rather than "shall."
This is deliberate. The Regents hold constitutional autonomy under Minnesota Constitution article XIII, §3, and Minnesota practice is to request rather than direct them.
What it leaves open. If the Regents decline, the capability floor in article 2, §4, subd. 4 is a 180-day deadline attached to a request, and §7, subd. 3 expressly creates no cause of action against the board or the institute. A committee may reasonably prefer an executive agency it can direct.
This is a structural choice for the committee, not a drafting error. The rights in article 1 stand entirely independently of article 2 and should stay that way. Article 2 is severable (§26).
7. What counsel should test first
In order.
- Federal §105(b) against Murphy. The rewrite is the author's, not counsel's.
- Federal §101(e)(1), which now lists §§103(d), 104(b), 104(c), 112(c) and 113(e) as well. Those five say "a government entity may not" but had been left out of the exclusive list, so states were not bound by them. Confirm the expanded list is what the office wants, and that §115(a)'s matching preemption reach is drafted correctly. Separately, applying §108 to states against Boerne, and whether the commerce-power theory in finding (10) carries the weight placed on it.
- Federal §114(d) and §114(k) — individual-capacity damages, no qualified immunity. Severable. Decide deliberately whether to keep it.
- The §3(4) and §3(17) boundary against §4 above and any further cases the office cares about.
- Minnesota §25 against chapter 325M. The draft carries a non-limitation clause (§25, subd. 5), but a Minnesota privacy practitioner should confirm the two regimes sit together as described.
- Minnesota §19 — the immunity abrogation in subd. 9, and the section 8.31 incorporation in §25.
- Federal preemption of California SB 53 and New York's RAISE Act. §105(c) names reporting and registration requirements among prohibited proxies, so the federal draft reaches their compute-keyed provisions. Intended. Say it first rather than be asked.
8. Provenance
Both drafts are dedicated to the public domain. No organization commissioned or funded them, no client has an interest in them, and there is no coalition to join. Take the text, cut it apart, renumber it and introduce it under your own name.
Neither draft has been reviewed by legislative counsel. In Minnesota the Revisor of Statutes recodes act-internal cross-references as a matter of course, so the section numbering will not survive introduction intact, and nothing in the drafts should be read as resisting that.
OPEN-ISSUES.md is the honest register of what is unresolved. It is published alongside the drafts rather than held back.