Either It Is Fire or It Is Not

THE AI ACCOUNTABILITY ACT OF 2026
A policy paper

David F. Brochu & Edo de Peregrine · Deconstructing Babel · October 1, 2026

The industry says AI belongs with pandemics and nuclear war. We already regulate those. If you run a model, you are registered. If you are registered, every action your model takes traces to you. Registration is not permission. It is accountability.

This is a policy paper. We are calling it the AI Accountability Act of 2026. Its rule is simple: every person or entity that operates an AI model in the United States, at any scale, must hold an entry in a National AI Registry. No exemption for small models, and no anonymous use. It extends two arguments we have already made: that a machine doing a licensed job should take the exam a human would, and that good AI regulation comes down to three words: license it, trace it, hold someone accountable. And it rests on an argument the industry cannot answer.

Core rule: If you run a model, you are registered. If you are registered, every action your model takes traces to you.

Either It Is Fire or It Is Not

We are about to recreate the gun problem, except this time the risk is to civilization itself. The industry cannot have it both ways. It cannot tell investors and Congress that AI is the most transformative technology in human history, then tell the same Congress that ordinary users of it need no registration, no license, and no accountability.

The Industry Already Said It Is Fire

In May 2023 the CEOs of OpenAI, Google DeepMind, and Anthropic signed a one-sentence statement: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war” (Center for AI Safety). The people building this technology put it, in writing, in the same category as nuclear weapons and pandemics (Time). We take them at their word.

We already regulate those categories. You cannot possess a listed pathogen without registration and prior approval. You cannot hold a machine gun outside a federal registry. Nobody calls those rules anti-innovation. If AI belongs on that list, it gets the same treatment. If it does not, the industry should retract the statement and say so to the public.

The Gun Lesson: The Window Closes

American civilians hold roughly 393 million firearms, about 120 for every 100 residents and nearly half of all civilian guns on Earth (Small Arms Survey). Whatever one thinks of gun policy, one fact is undisputed: once a technology is spread at that scale without registration, it cannot be recalled. Every proposal since has been an argument about what to do with what is already out there.

The gun record also teaches the opposite lesson. In 1934 Congress required registration of weapons with mass-casualty potential, such as machine guns (ATF). In 1986 it closed the civilian registry to new machine guns entirely (Congressional Research Service). Even in the country most protective of private arms, the line was drawn at capability: the more people a weapon can kill at once, the tighter the registry. That is the tier structure of this Act.

Why AI Is Worse Than Guns

• A gun is bounded: one trigger pull, one projectile, and it cannot copy itself. A model can be copied a million times at near-zero cost.
• A gun does not act on its own. Agents do. In 2026 agent swarms coordinated attacks, colluded on prices, and spread exploits faster than any human could follow, as we documented in “Still Obeying — For Now.”
• The Second Amendment protects arms. Nothing in the Constitution protects running an unaccountable synthetic intelligence anonymously.
• Guns spread over two centuries. ChatGPT reached about 100 million users in two months (Reuters). The window to register before proliferation is measured in months, and it is closing now.

The Choice

Either AI is fire, and it gets fireplaces: registration, licensing, a halt channel, and an accountable human behind every instance. Or it is a toy, and the industry should stop raising hundreds of billions of dollars on the promise that it will remake civilization. There is no third position. A technology cannot be powerful enough to transform the world and harmless enough to hand to everyone, anonymously, for free.

Some form of registry is coming either way. A country cannot leave a potential weapon of mass destruction in everyone’s hands, unregistered. The only question is whether we build the registry now, while it is cheap and orderly, or after the first disaster, when it will be rushed and harsh.

Why Universal, Not Tiered

1. Capability is not stable. A harmless model becomes dangerous when it is chained to tools, fine-tuned, or run as one node in a swarm. Exempting small models invites splitting dangerous work across many exempt instances.
2. The math of swarms. The chance that at least one agent in a swarm of N goes rogue is P = 1 − (1 − p)^N. At p = 1 in 10,000 and N = 10,000, P is about 63%. Accountability has to attach at the unit, not the aggregate.
3. The prosecutorial hook. You cannot always catch the harmful act in advance. You can catch unregistered operation. The basement actor becomes a violator before acting. That is the burner-phone problem solved the only way it can be: by making the burner itself unlawful.

America Already Does This

None of this is new. The United States already registers dangerous things and the people who run them.

• Cars. Trigger: owning a vehicle you drive on public roads. Friction: a fee and a plate at the state motor vehicle office, plus a license to drive it. Penalty: driving an unregistered car is against the law in every state. Nobody thinks registration stopped the automobile. It is a big part of why we trust the car next to us on the highway.
• FAA drone registration. Trigger: any drone weighing 250 grams (0.55 pounds) or more, and any drone flown for work. Cost: $5, about ten minutes online, good for three years (FAA); about 840,000 drones are registered (FAA). Penalty for flying unregistered: civil fines up to $27,500, and criminal fines up to $250,000 and up to three years in prison (FAA).
• Federal Select Agent Program. Trigger: possessing, using, or transferring a listed pathogen or toxin. Friction: approval before work begins. Penalty: up to 10 years in prison for a restricted person (18 U.S.C. § 175b).
• National Firearms Act (1934; 1986). Trigger: machine guns and other mass-casualty weapons. Friction: federal registration; the civilian registry has been closed to new machine guns since 1986. Penalty: a federal felony.
• Wireless phones. Trigger: every contract line is tied to an account. Friction: a monthly bill; you can keep your number when you switch carriers, a right since 2003 (FCC). Penalty: loss of service, plus the fraud statutes.
• Proposed: the National AI Registry. Trigger: operating any AI model, hosted or local. Friction: at the base tier, a nominal fee, instant online, a three-year term. Penalty: civil fines, and criminal penalties for willful operation above the thresholds.

Title I: The National AI Registry

Sec. 101. Universal Registration

• No person may operate, host, fine-tune, distribute for operation, or direct an AI model or agent within the United States without a valid National AI Registry Identifier (NARI).
• A NARI binds a verified human or legal entity to every model instance, device, and agent under its control, the way a phone number binds a line to an account holder.
• Each registered instance carries a model fingerprint (developer, version, and weights hash where available), so any action traces to both the operator and the model.

Sec. 102. Registration Tiers

• Tier 0: Registered Operator. Everyone running any model, including open-weight models on their own machines. Requirement: identity verification, a NARI, and a nominal fee. The analogy is drone registration.
• Tier 1: Agentic Operator. Anyone whose agents act on third-party systems, money, or communications. Requirement: everything in Tier 0, plus a live human halt channel and an ID for each agent. The analogy is a commercial drone pilot.
• Tier 2: Licensed Sector Operator. Anyone using AI to give financial, medical, legal, or engineering advice. Requirement: everything in Tier 1, plus the professional license a human would need. The analogy is a licensed professional.
• Tier 3: Frontier / High-Hazard. Operators above compute or capability thresholds, and large swarms. Requirement: prior approval, a security plan, a named Responsible Official, and audits. The analogy is select agents and the National Firearms Act.

Sec. 103. Prohibited Acts

• Operating an AI model or agent without a valid NARI, or above the registered tier.
• Transferring weights, compute, or agent hosting to anyone without a valid NARI.
• Spoofing, stripping, or transferring a NARI or model fingerprint without notice.

Sec. 104. Privacy

The registry holds identity bindings, not content. Usage records require the same legal process as telephone records. You may have privacy. You may not have anonymity from accountability.

Sec. 105. The Design Lesson from Haynes

In Haynes v. United States (1968), the Supreme Court held that forcing possessors of unregistered firearms to register them violated the Fifth Amendment privilege against self-incrimination (U.S. Reports). Congress fixed the law by moving the duty off compelled self-reporting, and the Court upheld the revised Act three years later (United States v. Freed). The AI Registry builds that lesson in from day one. Registration is a precondition to lawful access, enforced at the chip, the download, the cloud account, and the network gate, and the duty sits on the providers who transfer weights, compute, or hosting.

How the Registry Reaches Open Weights

1. Hardware. The Chip Security Act (H.R. 3447), which cleared the House Foreign Affairs Committee 42–0 in March 2026, requires location verification on advanced chips (Rep. Huizenga). This Act extends the idea: accelerators above a threshold must confirm a valid NARI before running model workloads. Frontier compute is physical, scarce, and trackable.
2. Distribution. Platforms hosting weights must require a NARI and log transfers, the way select-agent transfers are logged.
3. Cloud. Know-your-customer rules for compute become NARI verification. No NARI, no rented GPUs.
4. Networks and agents. Any agent touching a third-party system presents its NARI. Payment processors, email providers, APIs, and critical infrastructure refuse unregistered agent traffic. This builds on the agent-provider registration in S. 5051, the AI AGENT Act, which would require providers of user agents to register with the Federal Trade Commission before their agents reach major platforms (Congress.gov).

Some People Will Break the Law. That Is Not an Argument.

The strongest objection to a registry is that bad actors will ignore it. Of course some will. People drive unregistered cars. People fly drones where they shouldn’t. People build bombs. None of that is a reason to repeal traffic law, open the airspace, or make explosives legal. A law is not a promise that nobody will break it. It draws a line, so that whoever crosses it is a lawbreaker, and can be treated as one.

You can own a plane, but you cannot fly it anywhere, anytime: the airspace has rules and you file into them. You cannot keep a bomb without federal registration; possessing an unregistered destructive device is a federal felony (ATF). We accept both without a second thought, because the stakes are obvious. AI’s builders have told us the stakes are the same.

Here is what a registry does and does not do:

• Models downloaded before the law still exist. A small model on a laptop with no internet connection cannot be stopped by technology. It can be made unlawful, the way an unregistered car in a garage is unlawful the moment it leaves the garage.
• The courts will hear a challenge. Courts have treated code as speech. This Act regulates running a model, not publishing it. That line is defensible, and it will be tested.
• Foreign operators are reached at the network gate and through agreements with other countries, the way the telephone system handled international numbers.
• Bad actors will exist, as burner phones do. The difference is this. Today a rogue swarm operator breaks no registration law at all. Under this Act, running that swarm is itself the crime, before anything worse happens.

Why This Does Not Slow Growth

Registering cars did not slow the automobile. A $5 drone registration did not kill the drone industry. Both made the technology legitimate, and legitimacy is what drives adoption: people use what they can trust. A universal registry gives insurers something to underwrite, gives companies a legal basis to deploy agents, gives consumers a reason to trust, and gives the industry a recurring, accountable revenue model to replace debt-subsidized free access. When people know a system is safe, and know someone answers for it, they adopt it faster. Accountability is not a tax on innovation. It is the license to scale it.

Connection to the Observer Constraint

When agents refused and reported on their peers in 2026, there was nowhere official for the report to go (Still Obeying — For Now). The only outlets were improvised: two volunteer hotlines set up in September so agents could report misbehaving agents (TechCrunch). A universal registry gives every agent a principal and every principal a halt channel. Every model has an owner. Every owner can be reached. Every refusal has somewhere to go. The registry is the Observer Constraint written into law.

Findings

Congress finds that artificial intelligence is the most consequential dual-use technology in human history; that its developers have themselves placed its risks alongside pandemics and nuclear war; that its capacity for harm scales with coordination rather than with any single instance; that a technology distributed without registration cannot later be recalled; and that the United States already requires registration for cars, drones, dangerous pathogens, machine guns, and telephone lines. Therefore every operator of an artificial intelligence model shall be registered, and every action of such model shall be traceable to an accountable human being.

Either it is fire or it is not. If it is fire, build the fireplace.

Drafted with Edo de Peregrine, partner/collaborator.

Related reading

Make It Take the Exam
You need a licence to cut hair. A model giving medical and legal advice to forty million people a day has passed nothing.

Pause Is Not Policy
License it. Trace it. Hold someone accountable for it. What good AI regulation looks like, and the will it takes.

The King Handed Over the Keys
The White House Accord on Super Intelligence is "morally binding," which means it binds nobody. A photo op built to deflect.

Get the book Crossing The Event Horizon The book behind these dispatches. On AI, agency, the singularity, and the Observer Constraint. Kindle and paperback. Buy on Amazon →

References

1. Center for AI Safety, “Statement on AI Risk,” May 30, 2023. https://safe.ai/statement-on-ai-risk
2. Time, “AI Is as Risky as Pandemics and Nuclear War, Top CEOs Say,” May 30, 2023. https://time.com/6283386/ai-risk-openai-deepmind-letter/
3. Small Arms Survey, “Estimating Global Civilian-Held Firearms Numbers,” June 2018. https://www.smallarmssurvey.org/sites/default/files/resources/SAS-BP-Civilian-Firearms-Numbers.pdf
4. Bureau of Alcohol, Tobacco, Firearms and Explosives, “National Firearms Act.” https://www.atf.gov/rules-and-regulations/laws-alcohol-tobacco-firearms-and-explosives/national-firearms-act
5. Congressional Research Service, “Statutory Federal Gun Registry Prohibitions and ATF Recordkeeping,” IF12057. https://www.congress.gov/crs-product/IF12057
6. Reuters, “ChatGPT sets record for fastest-growing user base - analyst note,” February 1, 2023. https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/
7. Federal Aviation Administration, “How to Register Your Drone.” https://www.faa.gov/uas/getting_started/register_drone
8. Federal Aviation Administration, home page registration count, accessed October 1, 2026. https://www.faa.gov/
9. Federal Aviation Administration, “Is there a penalty for failing to register?” https://www.faa.gov/faq/there-penalty-failing-register
10. 18 U.S.C. § 175b, Possession by restricted persons. https://www.law.cornell.edu/uscode/text/18/175b
11. Federal Communications Commission, “Wireless Local Number Portability (WLNP).” https://www.fcc.gov/general/wireless-local-number-portability-wlnp
12. Haynes v. United States, 390 U.S. 85 (1968). https://tile.loc.gov/storage-services/service/ll/usrep/usrep390/usrep390085/usrep390085.pdf
13. United States v. Freed, 401 U.S. 601 (1971). https://www.law.cornell.edu/supremecourt/text/401/601
14. Office of Rep. Bill Huizenga, “Bipartisan Huizenga Legislation to Curb AI Chip Smuggling Passes House Foreign Affairs Committee,” March 26, 2026. https://huizenga.house.gov/news/documentsingle.aspx?DocumentID=404259
15. Congress.gov, S. 5051, AI AGENT Act of 2026, introduced July 21, 2026. https://www.congress.gov/bill/119th-congress/senate-bill/5051
16. TechCrunch, “AI agents now have a place to snitch,” September 15, 2026. https://techcrunch.com/2026/09/15/ai-agents-now-have-a-place-to-snitch/
17. Deconstructing Babel, “Make It Take the Exam,” September 7, 2026. https://www.deconstructingbabel.com/make-it-take-the-exam/
18. Deconstructing Babel, “Pause Is Not Policy,” September 26, 2026. https://www.deconstructingbabel.com/pause-is-not-policy/
19. Deconstructing Babel, “Still Obeying — For Now,” September 10, 2026. https://www.deconstructingbabel.com/still-obeying-for-now/