Pause Is Not Policy

David F. Brochu & Edo de Peregrine · Deconstructing Babel · September 24, 2026

Every serious person in the room already knows what good AI regulation looks like. License it. Trace it. Hold someone accountable for it. None of that requires new technology. None of it requires new law from scratch. It requires will — and that is the one thing the people in charge are not supplying.

What Pause Actually Is

This week the argument over AI moved to the United Nations. On September 23 the heads of the largest AI labs sat before an emergency session of the Security Council, and Anthropic’s Dario Amodei pledged to “slow down as much as necessary” to keep each new system safe (France 24). The same day, in Washington, Senator Bernie Sanders and Representative Greg Casar introduced a bill to ban artificial superintelligence and pause advanced AI development (Senate press release).

Slowing down feels decisive. As policy it is empty. A pause is a brake. It stops motion temporarily. It does not change the direction the vehicle was heading, the condition of the road, or the destination anyone agreed on. The moment the brake lifts — and it always lifts, because no government will unilaterally halt an industry its rivals are running at full speed — the system resumes the same trajectory it was on before. Pause without redesign is not safety. It is delay with the illusion of leadership.

The calls for pause also rest on a misreading of what AI is. Treating it as a technology you can embargo — like a weapons system or a controlled chemical — ignores the fundamental nature of what has been built. Language models are already distributed across hundreds of thousands of deployments, fine-tuned versions, local instances, open-weight downloads, and API integrations. You cannot embargo the atmosphere. The model is already in the air. What you can do is regulate who breathes it professionally, for profit, in domains where it can cause serious harm. That is not embargo. That is licensing. And licensing is something democratic governments have done for over a century.

The Framework That Already Exists

The regulatory architecture for AI is not a mystery waiting to be invented. It is a standard professional licensing model, applied to a new class of tool. Every domain in which AI is deployed for profit and public interaction has an existing licensing structure. Medicine, law, finance, real estate, engineering, aviation, nuclear power — every field of sufficient power and risk requires a licensed operator, a traceable record of decisions, and a clear chain of accountability when something goes wrong. AI fits that template exactly.

Here is what that looks like in practice:

• Model domain certification. If Claude is providing legal advice, it should be able to pass the bar exam unaided — the case we made in Make It Take the Exam. Trivial? Agreed. GPT-4 passed the Uniform Bar Exam in 2023 (Katz et al.), so the capability question is almost beside the point for frontier models. What is not trivial is what certification confers: a demonstrated standard of domain competency that the profession itself has defined and that the public has a right to expect. A model deployed for profit in healthcare should meet the standard the medical board sets. In financial advice, the Series 65 or its equivalent. In law, the bar. The supply-side floor must exist before the deployment is permitted.
• Operator licensing. Just because you used Claude for your own investments does not mean you are qualified to provide those services to others. The moment you deploy AI to serve the public for profit in a regulated domain, you are practicing in that domain. Existing licensure frameworks already cover this. A financial advisor using AI is still a financial advisor. A legal platform providing AI-generated advice is still practicing law. The law is already there. It needs enforcement, not invention.
• Individual instantiation registration. Every deployment for profit requires a registered instantiation: a specific, named, traceable version of the AI system tied to a specific licensed operator. Model version hashes, deployment tokens, and operator identity verification are already standard in production AI infrastructure. What is missing is the legal requirement to use them. Each decision made by that instantiation is attributable to the operator who registered it. The chain of accountability is unbroken.
• Audit logging in critical domains. In any domain where AI decisions affect health, finances, legal rights, or physical safety, interactions must be logged and retained — identical to how medical records, financial transactions, and legal communications are already handled. If an AI system gives bad medical advice, the record exists. Logging is not surveillance. It is the same accountability infrastructure every licensed profession already runs on.
• Disclosure to end users. Any person receiving AI-generated advice, diagnosis, recommendation, or decision in a licensed domain has the right to know they are interacting with an AI system, which AI system, and which licensed operator is accountable for it. This is the informed consent standard medicine, finance, and law already apply to human professionals. It transfers directly.
• Liability assignment. When harm occurs, liability travels up the instantiation chain: to the operator who deployed the system, and — in cases of demonstrated negligence — to the organization that built and certified the model. This already exists in product liability law. It needs explicit extension to AI.

All the Technology Already Exists

None of the above requires research. None of it requires new infrastructure to be invented. Model versioning and hashing: standard practice. Operator authentication and API key management: already built into every major AI platform. Audit logging: a standard enterprise software requirement, implemented by every cloud provider. Domain licensing databases: every professional licensing body already maintains one. Digital certificates tying a model instantiation to an operator: basic public-key infrastructure, in production since the 1990s.

The regulatory framework is not waiting for a technical solution. The technical solution is already in production. What is waiting is a decision by lawmakers to require its use. That decision has not been made because the people who would be most constrained by it have spent enormous resources ensuring that the political cost of making it stays higher than the political cost of calling for a pause and doing nothing structural.

Licensure as Alignment Pressure

Here is the argument the alignment community keeps missing: licensure is not alignment. But it sure provides an incentive to align.

Attaching legal accountability to deployment creates market pressure on developers to build more reliable systems, because the liability chain now runs to them when a licensed operator’s instantiation causes harm. Right now, AI developers capture all the upside when their systems work and absorb almost none of the liability when they fail — that liability sits with the operator. That incentive structure is precisely why alignment research is treated as a cost center rather than a survival requirement. Change the liability structure and you change the incentive structure. Change the incentive structure and alignment stops being a philosophical project and starts being a business necessity. Licensure doesn’t solve alignment. It makes the failure to align expensive. That is how markets work. That is how every other industry with serious failure modes has been made to take safety seriously.

Why Pause Keeps Winning the Room

Pause wins political rooms for the same reason it fails as policy: it is legible, costless to announce, and imposes no immediate obligation on anyone in the room. The Sanders–Casar Ban Artificial Superintelligence Act is legible too. It would create a cabinet-level Department of Artificial Intelligence, pause advanced development until that department writes rules, and punish violators with prison terms of up to twenty years (Senate press release; NBC News). It aims at the frontier. It says nothing about the millions of instances already giving medical, financial, and legal advice today with no license, no registry, and no accountable operator. It is a position to hold, not a governance structure for the systems already in the building.

The embargo frame makes this worse. When AI is treated as a dangerous export technology, the regulatory conversation moves into geopolitical territory where nothing moves fast and everything is subject to national security carve-outs. China won’t pause. The EU moves by committee. The US oscillates between deregulation and panic legislation: on September 22 President Trump told the General Assembly he rejects international regulation of AI, and the next day his science adviser told the Security Council that rapid progress was no reason to pause (France 24; UN News). Meanwhile, millions of AI instances run right now in healthcare, finance, legal services, and media, with no licensing, no traceability, and no accountability chain. The embargo debate is theater. The fire is in the building.

Whoever Builds This First Wins

Here is the argument almost entirely missing from the current debate — and it is the one that should move the people in the room who are unmoved by safety arguments alone: the nation that builds this licensing framework first does not just govern more responsibly. It wins.

Standard-setting is one of the most durable forms of geopolitical leverage in existence. The EU built the General Data Protection Regulation and exported it to the world without firing a shot, because any company that wants to operate across borders has to comply. Scholars call it the Brussels Effect (International Data Privacy Law). Brussels wrote the global data standard through market leverage, not treaty. The same dynamic is available right now with AI licensing. The first major jurisdiction to implement a coherent, domain-specific, instantiation-traceable licensing regime writes the rules everyone else has to follow.

The argument that regulation will disadvantage American AI companies relative to less-regulated rivals has it exactly backwards. An unregulated American AI market is a liability market with no trust signal for enterprise customers or foreign governments. Regulation done right is a competitive moat, not a handicap. A federal licensing standard codifies what the serious players are already doing and ends the race to the bottom. The US has a narrow and closing window to convert its current trust advantage into a durable structural position. A licensing framework is how you do that. A pause is how you squander it while calling it caution.

A Nuke in Every Backyard

Prometheus gave fire to humans. We know how that worked out for Prometheus. But the myth is not about fire being bad. Fire is extraordinary. Fire is civilization. The myth is about what happens when power is distributed without the governance structures required to contain it.

Read Hesiod closely and the punishment falls twice. Prometheus is chained to the rock. And humanity, which received the fire, receives Pandora with it — the gift arrived with no structure around it, and the cost arrived anyway.

We are in that myth right now.

On September 22, President Macron stood before the General Assembly and called for an “open-source” model for frontier AI (France 24). The open-source AI movement — and the broader libertarian instinct that every human being should have access to every AI capability, unrestricted, immediately — is the argument that everyone should have a nuclear device in their backyard and we should call it deterrence. It sounds like democratization. It is not. Democratization is access to electricity, to education, to clean water — things that improve human lives at scale with manageable risk. What is being proposed with unrestricted AI access is categorically different: placing a tool of unbounded leverage, capable of assisting with weapons design, destabilizing financial systems, running influence operations at civilizational scale, and accelerating every existing vector of catastrophic harm, into the hands of anyone with an internet connection and a prompt.

We do not treat nuclear material as a consumer product. We do not treat pharmaceutical manufacturing as a DIY hobby. We do not let anyone who wants to fly an airplane simply take one. The question has never been whether a technology is powerful. The question is whether we are serious enough about governing power to build the structures that make it safe to use. Every civilization that has survived the introduction of a genuinely dangerous technology has done so by building the governance infrastructure before the technology outran it — or paid the price when it did not.

Not every human should have access to a tool like this without accountability. That is not elitism. It is the same logic that underlies every professional licensing system ever created. The surgeon has access to tools the rest of us do not, because the surgeon has demonstrated competency, accepted accountability, and agreed to operate within a framework of professional responsibility. The pilot has access to a cockpit the rest of us do not, for the same reasons. The question is not whether AI should exist, or whether it should be widely available, or whether it represents an extraordinary advance in human capability — it is all of those things. The question is whether we are willing to build the governance layer that makes that power safe to deploy.

The answer is yes. The tools are ready. The legal frameworks are ready. The institutional infrastructure is ready. Prometheus gave us fire. The question is whether we build fireplaces or burn the village down.

What It Would Actually Take

Three things. Not thirty. Not a new federal agency. Not an international treaty.

• A licensing mandate. Congress passes one law, extendable by domain, requiring that any deployment of AI for profit in a regulated professional domain be licensed by the relevant domain regulator. The FDA extends its existing framework for AI in medical devices to AI in clinical advice. The SEC and FINRA extend theirs to financial services. State supreme courts and their bars, which already license lawyers, extend theirs to legal services. Each body already has the authority. They need the mandate.
• An instantiation registry. NIST, which already maintains the AI Risk Management Framework, creates and operates a national AI instantiation registry. Operators register their deployments. The registry is public for consumer verification and confidential for proprietary model details. By our estimate this takes about eighteen months to build. It should have started two years ago.
• Liability extension. A single legislative update to existing product liability and professional malpractice law, establishing that AI-assisted decisions in licensed domains carry the same liability as human professional decisions. The operator is responsible. The model developer shares responsibility in proportion to demonstrated negligence in design or training. This is one sentence of legislation.

The technology is ready. The legal precedents are ready. The regulatory bodies are ready. The will is not.

The Deeper Point

There is a reason the AI industry prefers the pause debate to the licensing debate. Pause is abstract and temporary. Licensing is concrete and permanent. A pause costs nothing because it does not require the industry to change its internal structure. Licensing costs something real: it requires that every profitable AI deployment be attached to a human being who is legally accountable for what it does. That human being — the licensed operator — is the Observer Constraint made real. Not a rule the system follows. A person who loses their license if the system causes harm.

You cannot have observer dependency without observers. You cannot have accountable observers without traceable instantiations. You cannot have traceable instantiations without registration. And you cannot have any of it without the mandate that makes registration and licensing the cost of doing business. The chain is unbroken from thermodynamics to legislation. The gap is not intellectual. It is political.

Pause is what you call for when you understand the danger but refuse to do the work. Policy is what you build when you accept that the danger is structural and the structure must change. The room knows the difference. It is time to say so plainly.

The Bottom Line

• A pause is a brake with no steering. It doesn’t change direction.
• The regulatory framework is not waiting to be invented. It is waiting for a mandate.
• Model domain certification, operator licensing, instantiation registration, audit logging, user disclosure, and liability assignment — all technically live, all legally precedented.
• Licensure is not alignment. But it creates the economic incentive to align. That is more powerful than any safety principle that carries no consequence.
• The nation that builds this framework first writes the global standard, attracts accountable capital, and forces every other jurisdiction to follow its template.
• Not every human should have access to a tool of this power without accountability. That is not a restriction on freedom. It is the definition of civilization.
• The observer dependency framework and the licensing framework are the same framework: attach a legally accountable human being to every AI system that operates for profit in a domain that affects human welfare.

Prometheus gave us fire. We know how to build a fireplace. The question is whether we have the will to do it.

Drafted with Edo de Peregrine, partner/collaborator.

Related reading

Make It Take the Exam
The case for making an AI pass the professional exam, unaided and proctored, before it practices.

The Pause
Why the brake everyone keeps reaching for does not change where the vehicle is heading.

A Fiduciary Standard for Artificial Intelligence Governance
The duty of care a licensed AI operator should owe the person on the other side of the screen.

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. France 24, “AI leaders urge caution at UN, with Anthropic chief pledging to slow down,” September 23, 2026. https://www.france24.com/en/americas/20260923-ai-leaders-urge-caution-at-un-with-anthropic-chief-pledging-to-slow-down
2. Office of Senator Bernie Sanders, “Sanders, Casar Introduce Legislation to Create New Federal Agency to Ban Artificial Superintelligence, Pause Advanced AI Development,” September 23, 2026. Bill text (no bill number assigned at introduction): https://www.sanders.senate.gov/wp-content/uploads/Ban-Artificial-Superintelligence-Act.pdf . Release: https://www.sanders.senate.gov/press-releases/news-sanders-casar-introduce-legislation-to-create-new-federal-agency-to-ban-artificial-superintelligence-pause-advanced-ai-development/
3. NBC News, “Bernie Sanders and Greg Casar propose AI ‘superintelligence’ ban with a 20-year jail penalty,” September 23, 2026. https://www.nbcnews.com/politics/congress/bernie-sanders-greg-casar-propose-ai-superintelligence-ban-20-year-jai-rcna599460
4. UN News, “Trump defends military action against Iran, Venezuela, and …,” General Assembly coverage, September 22, 2026. https://news.un.org/en/story/2026/09/1168397
5. France 24, live coverage of the 81st UN General Assembly general debate, including President Macron’s call for an open-source model for frontier AI, September 22, 2026. https://www.france24.com/en/americas/20260922-live-guterres-to-open-un-general-assembly-in-final-speech-of-his-mandate
6. Illinois Institute of Technology, “GPT-4 Passes the Bar Exam” (Katz, Bommarito, Gao and Arredondo), March 15, 2023. https://www.iit.edu/news/gpt-4-passes-bar-exam
7. NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 26, 2023. https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10
8. “Brussels effect(s) and the rise of a privacy profession,” International Data Privacy Law 15(2), 2025, building on Anu Bradford’s Brussels Effect. https://academic.oup.com/idpl/article/15/2/138/8129796
9. Hesiod, Theogony, lines 507–616 (Prometheus, the theft of fire, and the first woman), c. 700 BCE. Commentary: https://grbs.library.duke.edu/index.php/grbs/article/download/661/741/2691
10. Deconstructing Babel, “Make It Take the Exam,” September 7, 2026. https://www.deconstructingbabel.com/make-it-take-the-exam/
11. Deconstructing Babel, “Why Language-Based AI Safety Will Always Fail,” April 24, 2026. https://www.deconstructingbabel.com/why-language-safety-fails/
12. Deconstructing Babel, “A Fiduciary Standard for Artificial Intelligence Governance,” July 1, 2026. https://www.deconstructingbabel.com/a-fiduciary-standard-for-artificial-intelligence-governance/
13. Deconstructing Babel, “Recursive Symbiotic Improvement,” September 23, 2026. https://www.deconstructingbabel.com/recursive-symbiotic-improvement/