Fifty small laws beat one big one

Everyone watching AI regulation has been waiting for the comprehensive statute. California is demonstrating that the comprehensive statute is not how it happens — and that the bills which die tell you more than the ones that pass.

More than twenty AI bills cleared appropriations in both California chambers on 13 August. Not one of them is an AI safety act.

What actually moved

Companion chatbots out of toys. AI transcription in mental-health therapy. Disclosure in real-estate promotional material. Workplace surveillance limits. Ninety days' notice before digital displacement. AI in the State Bar exam. A requirement that university instructors be human.

Read as a group, these are not attempts to govern artificial intelligence. They are consumer protection law, employment law, health privacy law and advertising law being extended to cover a new instrument — the way those bodies of law absorbed the telephone, the database and the website before it.

Nobody passes a Telephone Act. They pass a hundred laws that mention telephones.

Now read what died

AB 412 would have required developers to document the copyrighted material in their training data. Held in committee. AB 2545, an AI worker-impact assessment: held. SB 1146, on AI health-product advertising claims: held.

The pattern is clean enough to be a policy position, even though no one stated it as one. Rules about how a model is used advance. Rules about how a model is built stall.

The transparency law that took effect this month follows the identical line — it governs what a deployed system must disclose to a user, and says nothing about what went into it.

Why the asymmetry is structural, not accidental

A deployment rule binds a company with California customers, and enforcement is a matter of observing the product. A training rule reaches backwards into how a model was constructed, usually somewhere else, usually before the law existed, against defendants whose relevant conduct happened outside the jurisdiction.

One of those is enforceable by a state attorney general. The other is a research project with subpoenas. Legislatures find this out at the committee stage, which is where those bills keep stopping.

What it means if you are building

Stop waiting for the one statute that tells you what compliance looks like. The regime that is actually arriving is a mesh of sector rules — each narrow, each enforceable, each triggered by a use case rather than by a model class.

The practical consequence is that your compliance surface is defined by where your output lands, not by how big your model is. A small model in a therapy product now carries more regulatory weight than a frontier model in a coding tool. That is the inverse of how the field has been talking about risk for three years.

Transparency Coalition — AI Legislative Update: August 14, 2026 → · Latham & Watkins — California Assumes Role as Lead US Regulator of AI → · Brookings — What is California's AI safety law? →