California AB 1883 and your engagement dashboard

Two functions, and only two. AB 1883 was chaptered as Chapter 853 and adds Part 5.8, commencing with Section 1580, to Division 2 of the Labor Code. The operative sentence is Labor Code 1581(a): "An employer shall not use a workplace surveillance tool that uses artificial intelligence on employees for either of the following: (1) Recognizing, or making inferences or predictions about, an individual's emotional state. (2) Collecting neural data." Subdivision (b) then says the section does not prohibit an employer from using a workplace surveillance tool to ensure safety, or otherwise if it does not meet the specifications in subdivision (a). The bill carries no urgency clause and sets no operative date of its own, so it takes effect January 1, 2027. Neural data is defined at Section 1580(e) as information generated by measuring the activity of an employee's central or peripheral nervous system that is not inferred from nonneural information, which means an EEG headband and not a webcam. Enforcement runs through the Labor Commissioner or a public prosecutor, with a civil penalty of up to 500 dollars per violation and a private right of action. The definition of employer is wide: any person or governmental entity that directly or indirectly employs or exercises control over wages, hours or working conditions, including charter cities, the University of California, labor contractors and the entities that contract with them. There are narrow exemptions for employers developing aircraft for the national airspace and for national security, military, space or defense work. Almost nobody reading this is inside those.

Yes, and that is the less interesting half of the answer. Section 1580(g) defines a workplace surveillance tool as any system, application, instrument or device that collects or facilitates the collection of employee data, activities, communications, actions, biometrics or behaviors by means other than direct observation by a person, including video or audio surveillance, continuous incremental time-tracking tools, geolocation, electromagnetic tracking and photo-optical systems. There is no intent requirement in that definition. There is no carve-out for software whose main job is something else. The phrase "facilitates the collection" reaches the note-taking bot and the analytics integration as well as the platform itself. So your webinar platform, your LMS and your meeting recorder are all workplace surveillance tools the moment they are pointed at employees, and that is true of every vendor in the category, not just the one you happen to use. Being a surveillance tool is not the violation. The test has two steps and most compliance conversations only run the first one. Step one: is it a surveillance tool. Step two, the one that decides the outcome: is artificial intelligence being used in it to do one of the two prohibited things.

In its ordinary form, no, and the reason is worth holding onto because it is the line you will be drawing over and over. Zoom's Webinars and Events product ships an Event Engagement Score that rates attendees on a 0 to 10 scale based on actions like chat activity, Q&A participation and poll responses. Those are behaviors. The same is true of attendance duration, camera on or off state, quiz results, time spent in a breakout and whether somebody opened the workbook. Zoom also ships an AI-Driven Key Moments feature that analyses reactions, chat spikes and Q&A surges to identify which segments of a webinar drove the most activity, and that is still counting things people did. The presence of artificial intelligence does not make a metric prohibited. What matters is the output the model is being asked to produce. Counting is not inferring. A tool that reports a person typed four times in the chat is recording an action. A tool that reports a person seemed frustrated is making a claim about their internal state, and 1581(a)(1) names exactly that claim. The practical version: if the output is a count of something observable, you are outside the prohibition. If the output is an adjective about a person, look harder.

This is the exposure, and it sits inside the same vendor as the safe feature. Zoom's Quality Management product produces both an engagement score and a sentiment score, and Zoom's own documentation describes engagement as a rated score based on talk-time ratio, response time lag and frequency of speaker changes, and sentiment as a rated score derived from AI analysis of conversation transcripts, noting that sentiment scores do not account for vocal tone or volume and read transcript content only. Strip the product language away and a sentiment score is a number that represents how positive or negative a person was feeling, produced by a model, from that person's own words, inside a tool that collects employee communications. That is the function Labor Code 1581(a)(1) describes. The conversation intelligence category sits in the same place: platforms in that market advertise sentiment analysis that tracks engagement and emotional signals across calls on Zoom, Teams and Google Meet. One distinction worth keeping straight. Zoom Contact Center's sentiment analysis is scoped to the consumer in the conversation rather than the employee, and 1581 is about employees, so consumer sentiment is outside it. The agent talk metrics running alongside it are about the employee. Same screen, different answer, which is why an inventory beats a policy here.

The duty runs to the employer. The statute says an employer shall not use, and it does not create a vendor obligation or reach an outside facilitator who is not the participants' employer. If you are a training consultant running a cohort for a California client, you are not the regulated party. You are, however, perfectly capable of creating the record that makes your client the regulated party. Think about what your post-session deliverable contains. If your engagement report hands a client a sentiment reading, a mood summary or an AI-written line about how a named employee seemed during a difficult module, you have produced an inference about an individual employee's emotional state, generated by artificial intelligence, out of a tool that collected their communications, and you have filed it in their employer's records. The client did not decide to do that. You decided it for them. The fix is small and it makes your reporting better anyway: deliver attendance, participation counts, completion and assessment results against stated criteria. That is also the shape of record Senate Bill 513 pushes employers toward, since training records became personnel records and the useful ones show what somebody can now do rather than how they appeared while learning it.

Five things, and none of them takes a budget. First, inventory every tool that touches a live session: the meeting platform, the LMS, the recorder, the note-taker, the analytics layer and any integration that reads transcripts. Second, go feature by feature and sort each output into counting or inferring. Counts of actions stay. Anything that returns a word or a score describing a person's state goes on the review list. Third, turn off the affect features you are not using, which in most accounts is all of them, because sentiment scoring tends to arrive switched on inside a bundle nobody evaluated. Fourth, send one written question to each vendor: does any feature in our configuration produce an output that describes, scores or predicts an individual employee's emotional state, mood or sentiment. Written, because a procurement file beats a memory in January 2027. Fifth, rewrite your own definition of engagement in your reporting template so it names the behaviors you count, which protects you whether or not a regulator ever reads it. One caution on the safety carve-out in 1581(b). It is there for safety systems. It is not a wrapper you can put around an engagement analytics product by asserting that attention monitoring keeps people safe.

The honest gap, and you should know where it is before you rely on any of this. Section 1580 defines seven terms. Emotional state is not one of them. The law tells you precisely what neural data is and then prohibits inference about emotional state without saying what an emotional state is. That leaves the boundary between a sentiment score and an engagement score drawn, at the moment, by vendor marketing copy rather than by the statute, and a vendor that rebrands a mood feature as an energy reading has changed nothing about what the model does. There is no implementing regulation and no case law as of October 2026, so nobody can tell you with certainty where a transcript-derived sentiment score lands if it is ever litigated. My reading is that it lands inside the prohibition, because the statutory language reaches inferences and predictions and not only direct measurement, and a transcript-derived score is plainly an inference. That is a reading and not a holding. Verify against the chaptered text of AB 1883 at the California Legislative Information site, and if you are about to make a decision that costs money or changes a product, run it past an employment attorney licensed in California before you act on it.