California's next two AI bills regulate the buyer, not the builder, and both are still unsigned

Two AI bills that would bind California employers are sitting unsigned on the governor's desk, and the window closes September 30.

AB 1883(opens in a new tab) is the one to read first. It was presented to the governor on September 10 and would amend the Labor Code to bar employers from running workplace surveillance tools that infer an employee's emotional state, or that collect neural data, meaning signals measured from the central or peripheral nervous system. Penalties reach $500 per violation, with injunctive relief, punitive damages, and attorney's fees available on top, and both the Labor Commissioner and public prosecutors can bring actions. Note what separates this bill from the rest of the California package. The AI statutes signed on September 9 regulate model developers and the third parties who audit them. AB 1883 regulates deployers, which is to say the reader of this newsletter, and it reaches state agencies, the UC and CSU systems, counties, cities, school districts, and contractors.

SB 1000(opens in a new tab) has been on the desk since September 2 and points at the supply side instead. It would delete the one million monthly user threshold from the California AI Transparency Act, a threshold that today exempts most vendors from that law's provenance and disclosure duties. Remove it and a long list of smaller generative AI suppliers acquires obligations their contracts with you almost certainly do not mention.

Neither bill is law yet, and that is the actionable part. An inventory answers one question cheaply this month: does any tool already running in HR, contact center quality assurance, or productivity monitoring produce a score for sentiment, engagement, attention, or mood? Those features tend to arrive bundled inside something else rather than bought on purpose, which is exactly why they survive a policy review. Finding them before a signature costs a few hours. Finding them after costs a remediation project.

The second demand on the same budget is quieter. Microsoft shipped 974 CVEs in a single Patch Tuesday(opens in a new tab) on September 9, a record, and The Register argues that AI-assisted bug hunting is why(opens in a new tab), with researchers turning models loose on components nobody has examined in years. The tempting response is to point the same tools at remediation, and the evidence there is poor: a 1Password study found AI-generated patches fully resolved the flaw 26.0 percent of the time, while 53.9 percent either failed or introduced something new, and Veracode measured an average 56 percent security pass rate across more than 100 models.

So discovery is scaling and remediation is not. Put that next to two bills that move AI compliance onto the deployer, and both land on the same operations headcount in the same quarter. Governance written into a policy and governance that holds under load are separate budget lines. Most AI programs have funded only the first.