California just licensed my profession. Two other headlines want to license yours out of existence.

Three things happened in the same week, and the contrast between them is the most useful thing I have read all year about where AI regulation is going.

California signed a law that says AI auditing is now a profession with entry requirements. New York City banned AI in elementary and middle schools. And a bill landed in Congress that would ban the development of superintelligent AI outright, with prison as a penalty.

One of those is regulation. The other two are prohibition wearing regulation's clothes. The difference matters more than the politics of any of them.

Governor Newsom signed AB 1405 and SB 813 on September 9. They are usually reported as one thing. They are two, and the distinction is the whole story.

AB 1405 creates a registry. Beginning January 1, 2029, a person may not offer, sell, or conduct a covered AI audit in California unless they are registered with the Government Operations Agency. A covered audit is one that assesses the controls, processes, or systems of an AI system for compliance with state law. The obligation attaches to the audit, not to where the auditor sits, so a firm in Dallas auditing for California compliance is inside the definition.

SB 813 is separate and higher. It creates the independent verification organization, a designation the agency grants to auditors that demonstrate expertise in assessing AI risk. The agency has until January 1, 2028 to write those criteria. An auditor can be registered under AB 1405 without being designated under SB 813.

The conduct rules in AB 1405 are the part most coverage skips. A registered auditor may not audit its own work, meaning any system it materially designed, implemented, or operated. An individual may not seek employment with the auditee during the audit, or be assigned to audit a company that employed them in a related role in the previous twelve months. Records are kept for ten years. Every report must state its own limitations, including what was in scope and not assessed, and where the evidence had gaps.

If those rules sound familiar, that is because they are auditor independence rules, lifted almost intact from the accounting profession.

AB 1405 contains a carve-out that I have not seen mentioned in a single piece of coverage. A registered auditor holding a current California Board of Accountancy license is deemed to satisfy both the reporting requirements and the independence standards, provided the report addresses the required subject matter and the auditor follows the California Accountancy Act, the AICPA Code of Professional Conduct, and AICPA attestation standards.

Enforcement splits the same way. For most registrants, the agency investigates and may remove them from the registry. For a licensed CPA or firm in good standing, the agency refers the matter to the Board of Accountancy, which investigates under its own authority.

California looked at the problem of who can be trusted to audit an AI system and decided the answer was already written down, in the profession that has spent a century arguing about independence, evidence, and what an opinion is worth.

You have two years and change. The registry does not exist yet, the criteria are not written, and anyone telling you to register this quarter is selling something. What changes now is procurement.

From 2029, buying an AI audit from an unregistered provider means buying an audit that could not lawfully be sold. Ask providers whether they intend to register. Ask whether they carry an independence framework today or plan to build one. Ask whether they would fail their own conflict test, because the provision that stops an auditor assessing its own work eliminates a large amount of what is currently sold as AI assurance, where the firm that built the governance program also grades it.

That last one is the real reshaping. The market has been full of build-it-then-bless-it engagements. California just made that structure unlawful in its largest market, on a two-year fuse.

New York City's Department of Education banned AI in elementary and middle schools. The Ban Artificial Superintelligence Act would prohibit developing superintelligent AI, pause advanced AI until a new federal regulator sets guardrails, and attach penalties up to twenty years imprisonment.

I want to take both seriously before I disagree with them, because the instinct behind them is not stupid. A twelve-year-old who hands an assignment to a chatbot has not learned anything. A superintelligent system built carelessly by people racing each other is a genuine risk, and the people raising it are not cranks.

But the response in both cases is to remove the thing rather than govern it, and that is the move that does not work.

I have a twelve-year-old. He is learning to read more fluently, working through mathematics, and writing his first real code, and AI is a significant part of how all three are happening.

What strikes me is not the capability. It is the patience. It will explain the same concept six different ways without a trace of irritation. It never sighs. It never makes him feel slow for asking again. A child who is behind, or anxious, or simply needs the fourth explanation instead of the second, gets something from that which a classroom of thirty cannot reliably provide, however good the teacher is.

A ban does not remove AI from that child's life. It removes it from the one setting where an adult is present to teach him how to use it well. He will still use it at home, unsupervised, with no one showing him the difference between asking it to explain a proof and asking it to write his essay. The ban does not protect him. It moves his learning out of sight.

The right response is instruction, not prohibition. Teach children what these systems are, where they are confidently wrong, what they do with what you type into them, and when using one means you did not do the work. That is a curriculum problem. We have solved curriculum problems before.

The argument against a unilateral pause is not primarily economic. It is that a pause is not a pause. It is a decision about who continues.

Our adversaries will not stop. They have said so. A US moratorium does not slow the frontier, it relocates it, to laboratories with no congressional oversight, no FOIA, no plaintiff's bar, and no obligation to publish a system card. The safest version of a powerful technology is the one developed somewhere that can be watched, sued, subpoenaed, and voted out. Prohibition forfeits exactly that.

There is also an honest problem with the bill's own terms. It would ban developing a class of system that has no agreed technical definition, using a threshold no one can currently measure, enforced by an agency that does not exist. Laws like that do not stop frontier labs, which can afford lawyers. They stop the university group and the startup, and they concentrate the frontier among the handful of firms large enough to litigate.

Guardrails that assume development continues, because it will.

That means requirements aimed at what a system does, not at whether it may exist. Evaluation before deployment, with published results. Incident reporting with real consequences for staying silent. Independent assessment by people who did not build the thing, which is precisely what AB 1405 just built the scaffolding for. Liability that attaches to deploying a system you did not test, rather than to the act of research.

And it means an actual national plan, because the three pieces are one problem and Washington keeps treating them as three. Model development, semiconductor manufacturing, and data center capacity with the power to run it are a single supply chain. We are currently regulating the first, subsidising the second, and letting the third get blocked county by county, including the moratorium fight underway in Philadelphia this week. A country cannot lead in AI while its grid interconnection queue runs five years long.

That is the plan I want argued about. Not whether to permit the technology, which was decided by everyone else on earth continuing to build it, but what we require of it, who checks, and whether we build the physical capacity to run it here.

California, of all places, got the shape of this right in the same week two other institutions got it wrong. AB 1405 does not ban anything. It says that if you are going to tell the public an AI system is safe, you had better be independent, keep your evidence, and put your name on it.

That is what governing a technology looks like. The other two are what giving up on governing it looks like.