Consent, retrofitted: what California's four-front AI legislative wave is actually doing

California is not getting ahead of AI deployment. It is chasing it across four domains at once, and the shape of the chase tells you more about the next two years of enterprise compliance than any single bill does. The federal vacuum is being filled state by state, and the filling is reactive, not designed.

On 11 June the California legislature had five AI bills sitting in Senate committee: AB 1883 on workplace surveillance, AB 1979 and AB 2575 on healthcare AI consent and disclosure, AB 2392 on generative AI procurement across the community college, CSU, and UC systems, and SB 928 requiring CSU instructors to be human. SB 947, on automated decision systems in employment, is also in committee. Colorado got there first on healthcare: Governor Polis signed HB 1139 on 2 June. New York closed its session by passing five thematically linked bills — kids chatbot safety, AI training-data transparency, the FAIR News Act, a data-centre moratorium, and a surveillance-pricing ban — all now waiting on Governor Hochul's signature.¹

Federal legislation on any of these subject areas is not advancing.

The pattern is what matters, not the individual bills. Every domain California is legislating is one where AI is already operating inside California institutions. Workplace monitoring tools are deployed. Healthcare AI triages and drafts notes in hospitals today. CSU instructors are already piloting AI-assisted course delivery; UC procurement of generative AI tools is live. The bills are not anticipating a future deployment; they are retrofitting consent architecture onto systems that have already shipped.

This is the structural fact I want to sit with, because it changes what the legislation is for.

What "retrofit" actually means

Consent architecture, as the term gets used in privacy law, is the set of rules that determine when a person must be told a system is operating on them, what they must be told, and what they can refuse. GDPR built it for personal data. HIPAA built a narrower version for protected health information. The California bills are attempting to build it, in 2026, for AI systems that have already been integrated into workflows.

The retrofit problem is not unique to AI. Cookie consent under GDPR was a retrofit; the web had been running on third-party tracking for a decade before anyone was asked to click Accept All. The retrofit worked, in a narrow sense — disclosure happened, banners appeared. It did not meaningfully reduce tracking. The disclosure became theatre, and the underlying architecture continued.

The honest read on AB 1979, AB 2575, and AB 1883 is that they will produce disclosure. Patients will be told an AI system reviewed their imaging. Workers will be given thirty days' notice that monitoring tools collect their keystrokes. The deployments will continue. The bills are not designed to stop the deployments; they are designed to make them legible. That is a real policy outcome, and it is smaller than the headlines suggest.

The counter-case here is real and I want to name it. GDPR's effect on cookies was thin, but GDPR's effect on enterprise data handling was not — internal processes, vendor contracts, and breach disclosure obligations all moved. A consent-and-disclosure regime can change behaviour even when the user-facing layer is theatre, because compliance costs alter what gets built. If California's healthcare AI bills force hospitals to document every AI-assisted clinical decision, that documentation requirement is a real constraint on deployment, regardless of whether patients read the disclosure form.

So: retrofit consent is not nothing. It is just less than the role-preservation framing would suggest.

SB 928, and the only bill in the slate that tries to do something different

SB 928 is the outlier. It does not regulate process; it names a role and says that role must be held by a human. CSU instructors must be human. This is, to my knowledge, the first time a US state legislature has written occupational designation into AI law.

There are two readings of SB 928 and the evidence does not yet decide between them.

Reading one: precedent-setting floor. If SB 928 passes and survives challenge, it establishes that a legislature can name a role as off-limits to AI substitution. The implications extend well beyond CSU classrooms — public defenders, social workers, parole officers, clinical psychologists in state systems all become available targets for similar statutory protection. The bill becomes a template.

Reading two: symbolic gesture, hollowed in implementation. "Instructor must be human" leaves enormous room. An AI system that drafts the lectures, grades the assignments, designs the curriculum, and answers student questions through a teaching assistant interface, while a human "instructor" signs off, satisfies the statute. The role is preserved nominally; the work is not. This is how human-in-the-loop requirements have generally fared in automated credit decisions and clinical decision support: the human is in the loop, signing.

I think reading two is more likely, and I think it is more likely for an unglamorous reason. Enforcement of education mandates against public university systems is structurally weak. Who sues CSU for using too much AI in a classroom? The student paying tuition has no obvious cause of action. The faculty union might, but unions have generally treated AI as a bargaining issue, not a litigation one. The Attorney General has bigger targets.

That said, and this is where I update against my own prior, if SB 928 gets paired with a private right of action, or with a procurement-level enforcement mechanism through AB 2392, the calculus changes. The bill on its own is symbolic. The bill plus enforcement teeth is not.

New York's package versus California's slate

New York did something California did not. Five bills, passed in one session window, covering complementary angles: children's exposure to chatbots, transparency of training data, compensation for news content used in training, infrastructure siting, and pricing discrimination via surveillance. Whether or not Hochul signs all five, the package was designed as a package. The bills speak to each other.

The New York approach reads as a framework. The California approach reads as a response.

California's five bills do not speak to each other. AB 1883 on workplace surveillance and AB 2575 on healthcare AI consent share no definitional architecture. AB 2392 on procurement and SB 928 on human instructors could have been designed as a single education-AI framework; they were not. Each bill has its own sponsor, its own committee path, its own definitions of what counts as an AI system.

This matters because regulatory fragmentation has a cost, and the cost compounds when fragmentation exists within a single jurisdiction as well as across jurisdictions. A health system operating in California, Colorado, and New York already faces three different healthcare-AI consent regimes. Inside California alone, a university health system covered by AB 2575 (healthcare AI consent), AB 2392 (procurement), and SB 928 (instructor designation) faces three different definitional schemes for the same underlying technology.

The conventional read is that California leads on tech regulation. On AI, in 2026, that read is wrong. New York is producing more coherent legislative product. California is producing more legislative product.

What this means for enterprise deployment

For readers thinking about how to deploy AI inside organisations that touch California, Colorado, and New York, the operational picture is this.

Compliance will be governed by whichever state has the strictest applicable rule, because most enterprise systems cannot economically maintain per-state behavioural variants. This is the Brussels Effect, run domestically. The strictest applicable rule for healthcare AI is currently Colorado's HB 1139, now in force. For workplace surveillance, it will likely be California's AB 1883 or AB 1221 (the broader workplace surveillance bill referenced in the consensus reading), depending on which advances.² For training data transparency, it will be New York's bill if Hochul signs.

Documentation obligations are the binding cost, not the substantive restrictions. The bills mostly do not prohibit deployment; they require that deployment be documented, disclosed, and auditable. Build the audit trail once, design it to the strictest standard, and the compliance cost is bounded. Skip that work and the cost compounds across every new state regime.

The federal vacuum will not close on the timescale that matters. There is no federal AI legislation moving in either chamber on workplace surveillance, healthcare AI, educational AI, or training data transparency. Enterprise planning has to assume the state-by-state regime is the regime for the next several years.

And on SB 928 specifically: do not plan a CSU AI deployment strategy on the assumption it will not pass, and do not plan one on the assumption it will be aggressively enforced if it does. Plan one in which the human signs off and the AI does the work. That is what the bill, as written, permits.

What I am watching

Newsom's pen, mainly. He vetoed SB 1047 in 2024 on economic-impact grounds, and the California slate contains at least three bills with comparable industry-cost arguments available. Hochul's pen on the New York package, second — particularly the training-data transparency bill, which is the one with the largest direct cost to model developers. And whether any of the California healthcare bills get amended to align with Colorado HB 1139's definitions before passage, because that alignment, or its absence, will tell us whether anyone in Sacramento is thinking about fragmentation as a problem worth solving.

I suspect no one is. The bills are responsive, not designed. That is the story.

Footnotes

¶