California Bans Workplace Emotion Tracking and Sole-AI Firing

Governor Gavin Newsom signed a workplace AI package on September 30, 2026 — not a frontier-lab shutdown mandate.

The Governor's office framed the bills as worker protections: no sole-AI discipline or firing, more transparency when AI drives mass displacement, and limits on AI-powered workplace surveillance. CNBC led with SB 947, the No Robo Bosses Act, after Newsom vetoed an earlier version in October 2025.

If you ship HR copilots, sentiment analytics, interview emotion AI, or agentic performance tools used by California employers, product scope changes. If you train frontier models, this is a different stack from SB 813 / AB 1405 auditor licensing, Executive Order N-9-26's kill-switch study, and the August data-center infrastructure bills.

TL;DR — what actually changed?

What Newsom signed on September 30

The official list is wider than two bills. The worker cluster labor leaders named on the dais was SB 947, SB 951, AB 1331, and AB 1883.

Newsom's line from the release: "AI should expand opportunity — not come at the expense of workers and families." California Federation of Labor Unions president Lorena Gonzalez, speaking on those four bills, called them first-in-the-nation guardrails for human oversight, limited surveillance, and worker voice.

The same announcement also signed healthcare judgment bills, AI transparency updates, deepfake/likeness rules, gene-synthesis customer checks, and an executive-order gag about calling AI "Artificial Intelligence." Those are adjacent politics. The product problem for most explainx.ai readers is the employment stack.

Healthcare bills in the same pile (AB 1979, SB 503) keep licensed clinicians able to use professional judgment when clinical decision tools are in the loop, and push developers to take reasonable steps against known bias. That is closer to the Kaiser advice-nurse surveillance story than to HR SaaS — but the same lesson applies: do not encode empathy as a KPI.

AB 1883 — emotion inference and neural data (Jan 1, 2027)

Enrolled bill text on LegInfo adds Labor Code Part 5.8, starting at section 1580.

Workplace surveillance tool is broad: any system that collects or facilitates collection of employee data, activities, communications, actions, biometrics, or behaviors other than direct in-person observation. The digest lists video, audio, continuous time-tracking, geolocation, electromagnetic and photoelectronic tracking.

The prohibition is narrower than that definition. Section 1581(a) says an employer shall not use a workplace surveillance tool that uses artificial intelligence on employees for either of:

1. Recognizing, or making inferences or predictions about, an individual's emotional state
2. Collecting neural data

Neural data means information generated by measuring activity of the central or peripheral nervous system, not inferred from nonneural information. A webcam that guesses "frustrated" from a face is emotion inference. A headset that records EEG is neural collection. A badge that only logs door swipes is neither — unless you bolt an affect model onto the stream.

Section 1581(b) is the savings clause practitioners keep missing: the section does not prohibit using a workplace surveillance tool to ensure safety, or otherwise, if it does not meet the specifications in subdivision (a). Safety cameras, SOS badges, and occupancy sensors stay on the table. Stress-from-video wellness dashboards do not.

Section 1583 carves out aircraft development for the national airspace and products or services for national security, military, space, or defense, and only where the tool is reasonably necessary to comply with a federal statute, regulation, or binding federal contract.

Effective date, verified: the enrolled AB 1883 text we reviewed has no "operative on …" section. California statutes without a stated operative date take effect January 1 of the following year. Recaps that say January 1, 2027 match that rule. Treat the governor's September 30, 2026 package announcement as the signature event; the bill was enrolled September 3, 2026 (presented to the Governor). Some employment-law recaps dated the signature to September 3 — that is easy to confuse with enrollment. Use gov.ca.gov (Sep 30) plus LegInfo enrolled text if you need a cite.

Enforcement: Labor Commissioner or a public prosecutor. Penalty up to $500 per violation, plus injunctive relief, and the part does not preempt stricter local ordinances.

SB 947 — No Robo Bosses Act (July 1, 2027)

LegInfo enrolled text adds Labor Code Part 5.5.5. Section 1526.7 is explicit: "This part shall become operative on July 1, 2027." That matches the recaps. Do not collapse it with AB 1883's January date.

Automated decision system (ADS) means a computational process from machine learning, statistical modeling, data analytics, or AI that issues a simplified output (score, classification, recommendation) used to assist or replace human discretionary decisionmaking and that materially impacts people. The statute excludes spam filters, firewalls, antivirus, identity and access management, calculators, databases, and datasets.

The employment core is section 1522(b)–(c):

• An employer shall not rely solely on an ADS when making a disciplinary or termination decision.
• If the employer primarily relies on ADS output for that decision, it must direct a human to corroborate using data that produced the output or other relevant corroborating information. The statute's examples: supervisory evaluations, personnel files, work product, peer reviews, and witness interviews (which may include relevant online customer reviews).
• If the employer cannot corroborate, or the human reviewer concludes the output is inaccurate, incomplete, or misleading, the employer shall not use that ADS output for the discipline or termination decision.

"Primarily relies" is not defined. Chamber of Progress told Newsom that gap could chill ordinary tools. That criticism is real. Until regulations or case law fill it, vendors should assume a conservative reading: if a manager would not have fired the person without the model's score, you are in primary-reliance land.

Post-use notice (section 1524), not a pre-deployment banner. When the employer primarily relied on an ADS, the affected employee gets a written, plain-language, stand-alone notice at the time they are informed of the decision, in the language used for routine communications. It must say ADS was primarily used, that a human reviewed and corroborated, give contact information for a human, tell the worker they can request a description of their own data used by the ADS, and state anti-retaliation rights.

That is a product requirement: generate the notice, log the corroborating artifacts, name a reachable human. A Slack "looks good" reaction is not a personnel file.

Other 1522 limits matter for agentic HR: do not use an ADS to violate labor, OSHA, employment, or civil-rights law; do not infer protected status under Government Code § 12940; do not predict and take adverse action because a worker exercised legal rights.

Burden shift in court (1526.1(c)): once it is shown an ADS was used for a disciplinary or deactivation decision, the employer must show it did not primarily rely on the ADS or that it complied with 1522 and 1524. "Deactivation" in that enforcement sentence is leftover drafting next to an enrolled bill that otherwise talks discipline and termination — do not assume gig coverage from that one word.

Carve-outs: collective bargaining agreements that explicitly waive the part in clear terms and provide protection from algorithmic management; federal aircraft / national-security contracts (same shape as AB 1883); CCPA businesses still face CPPA automated-decision rules. Civil penalty: $500 per violation.

How this differs from the frontier-lab stack

California is running two AI policy machines. Mixing them is how vendors write the wrong compliance FAQ.

SB 53 does not tell you whether your interview product can score "confidence." EO N-9-26 does not tell you whether a copilot can auto-draft a termination letter from a productivity model. AB 1883 and SB 947 do.

CNBC notes Newsom's 2025 veto objected to unfocused advance notice on even innocuous tools. McNerney's 2026 rewrite removed that pre-notification and stripped gig-worker protections that rideshare companies fought. The signed law is narrower than the slogan, and still first-in-the-nation on sole-AI firing.

Illinois already requires notice for some employment AI uses. CNBC's comparison is right on the distinction: Illinois did not ban relying exclusively on automated systems for discipline and termination the way SB 947 does.

What this means if you ship HR or agentic workplace tools

Emotion and "sentiment" features. If the model outputs valence, stress, "engagement," micro-expression labels, or interview "soft skills" inferred from face or voice, that is 1581(a)(1) territory for a California employer customer. Gate it, disable it in the CA tenant, or delete the feature. Marketing copy that says "understand how your team feels" is a discovery exhibit.

Neural / wearable stacks. EEG, EMG, "focus" headbands, and any sensor that measures nervous-system activity rather than inferring from keyboard telemetry is 1581(a)(2). Do not hide it as wellness.

Safety vs affect. You can keep a warehouse collision-avoidance camera. You cannot put an affect head on that camera and sell "fatigue emotion scores" as a safety upsell unless counsel is unusually brave. The savings clause is function-based.

Discipline copilots. A tool that recommends "terminate" or auto-opens a PIP from ticket volume must force a human corroboration package: evals, files, work product, or interviews. Store the reviewer's conclusion. If they say the model is incomplete, block the adverse action on that output. Ship the 1524 notice template in the same workflow.

Agentic performance agents. An agent that pages a manager is fine. An agent that is the decision, with a human clicking "approve all," is the fact pattern SB 947 is written to catch. Design four-eyes as a product primitive, not a professional-services add-on.

Shadow tools. Workers will keep pasting transcripts into consumer chatbots, which is the shadow AI workplace privacy problem. These bills regulate employer use of surveillance and ADS, not every employee using ChatGPT on a phone. Employer-procured copilots still need the gates.

Eval design. If you score "empathy" in a call-center QA model, you are back in Goodhart's law — the Kaiser nurses already lived the pilot. AB 1883 makes the emotion-inference version a Labor Code problem, not just a union talking point.

California effect. National HR vendors will ship one CA-safe configuration rather than 50. If you sell to any employer with California staff, treat Jan 1 and July 1 2027 as release milestones, not legal trivia.

Penalty math. $500 per violation is small next to a wrongful-termination verdict, but per violation plus public-prosecutor actions and fee shifting is how these statutes get used. The bigger cost is discovery: model cards, prompts, and logs of who overrode the score.

Companion bills you should not ignore

AB 1331 generally bars using a workplace surveillance tool to monitor employees in a workplace bathroom, unless a court order says otherwise. Workers may leave badges and similar tools behind when they enter, with exceptions for identification/safety tools that meet stated conditions. Employers may still check one-time entry and exit. Same $500 penalty shape.

SB 951 amends Cal/WARN: when a covered mass layoff, relocation, or termination is caused in whole or in substantial part by an AI system or other automated technology replacing or automating jobs, the existing 60-day notice must include extra content (including job functions being automated). EDD publishes summaries; a legislative report on AI hiring effects is due on or before January 1, 2028. If your product is sold as "reduce headcount with agents," your customer's WARN packet is now part of the implementation.

The September 30 list also includes SB 574 (lawyers remain responsible for core legal work) and AB 2392 (public higher-ed AI training and procurement standards). Those are professional-conduct and campus-procurement stories, not HR-SaaS stories.

For the broader US/EU map, use explainx.ai's EU AI Act and US policy guide. For a running list of what "AI ban" headlines actually prohibited, see the 2026 AI ban scorecard.

What people are asking

Does this ban bossware? No. Time-tracking, GPS, and cameras can stay. The illegal layer is AI emotion inference, neural collection, bathroom surveillance, and ADS-only (or uncorroborated primary) discipline/termination.

Is a productivity score illegal? The score itself is not banned. Using it as the sole basis to fire someone is, after July 1, 2027. Using it as the primary basis without a human pulling a personnel file or equivalent is.

Do I need a California-resident employee to care? The employer definition is broad (including many public entities and labor contractors). If the employer is covered and the worker is employed in California, assume yes. Multi-state remote teams should not wait for the first test case.

What about interview emotion AI sold to recruiters? Hiring is not the same statutory hook as discipline or termination in SB 947's 1522(b). AB 1883 still hits emotion recognition on employees. Applicants vs employees is a counsel question; "we only score candidates" is a thin wall if the same model later scores staff. Do not ship affect scoring into California deployments as a default.

Will Washington preempt this? Unclear. A federal "No Robo Bosses" bill from Sens. Markey and Schatz was, in CNBC's telling, dead on arrival. State labor law has a long preemption fight ahead. Ship as if California remains in force.

Is $500 the real risk? It is the statutory civil penalty. Retaliation claims, injunctions, attorney's fees, and parallel CCPA automated-decision rules are the rest of the stack.

Honest limitations

• "Primarily relies" is undefined. Business groups will litigate it. Product teams should not wait for a regulation to add a corroboration step.
• Gig deactivation was bargained out of the 2026 SB 947 that landed on the Governor's desk, per CNBC. Platform companies should not celebrate too hard: local ordinances can still go further, and 1526.2 preserves stricter local rules.
• Advance notice of ADS in daily work is not in the signed SB 947. Workers learn at discipline/termination time. That is worse for workers than the 2025 bill and easier for IT.
• AB 1883 does not define "emotional state." Vendors will argue a "toxicity" classifier is not emotion. Expect that fight. If the output is a feeling-word, treat it as in scope.
• These are not a substitute for SB 53 safety disclosures or for whatever N-9-26 recommends on November 16.

Related on explainx.ai

• California AI audit laws: SB 813 and AB 1405 — auditor licensing, not workplace ADS
• Newsom's AI kill-switch executive order (N-9-26) — frontier-model study, not HR copilots
• California data-center AI infrastructure bills (August 2026) — grid and ratepayers, not emotion AI
• Kaiser nurses on AI workplace surveillance — the empathy-scoring fact pattern AB 1883 now addresses
• Shadow AI workplace privacy
• Specification gaming and Goodhart's law
• EU AI Act and US policy guide
• Every 2026 "AI ban" story: what actually got banned?

Primary sources: Governor Newsom, Sep 30, 2026 · CNBC on SB 947 · AB-1883 bill text · SB-947 bill text

Bill text, operative dates, and signing details are accurate as of October 1, 2026 against gov.ca.gov, CNBC, and California LegInfo enrolled language. Chaptered statutes, regulations, and the first enforcement actions may refine "emotional state" and "primarily relies." This is not legal advice.