On November 3, 2026, I’m presenting “Employee AI Monitoring” for the Future of Privacy Forum (event registration information). To coincide with that presentation, I thought it was a great time to publish this article on the same topic to get everyone ready to dive in!
California just drew the first real lines around AI and employee monitoring, and the rest of the country is right behind it. Here’s what changed, what it means for everyone who hires, manages, or counts on a paycheck, and why last year’s “AI void” argument only got stronger.
Last year I wrote an article for Reuters, “Stepping in the AI void in employment: Why state AI rules now matter more than federal policy” arguing that when it came to AI in employment, Washington had gone quiet and you had to look to the states.
A year later, while Washington did not stay quiet, it also did not move the needle on this issue. Rather, it issued an executive order, stood up a litigation task force at the Justice Department, published a legislative framework, and went into federal court to knock out a state AI law for the first time ever. That being said, not one state law got preempted.
Meanwhile the states kept legislating, the courts kept letting AI hiring cases through, and on the last day of September, California signed a package of workplace bills that moves the whole conversation from “who gets hired” to “who gets watched.”
The thing a lot of companies still have not come to terms with is that the AI-in-employment problem was never only a hiring problem. It is a monitoring problem. It starts before someone is hired, follows them through the workday, into the bathroom, home at night, and out the door when the layoff comes. And the law is now catching up at every one of those points.
I will skip a federal summary because there really is not one, other than deregulation. See Exhibit Order 14281 (April 2, 2025), which directed federal agencies to “eliminate the use of disparate-impact liability in all contexts to the maximum degree possible.” And while many business clients think the absence of federal regulation or deregulation is good for them, deregulation is not de-risking; it is re-routing. For every enforcement road that the federal government closed, a state agency or a plaintiff’s lawyer found a detour around it.
As the hiring cases got bigger, they stopped being only about bias. The one-to-watch case is Mobley v. Workday, which is now nationwide in discovery. I’ve written about this case before in “AI on Trial: Implications of the Workday Lawsuit for Automated Hiring”, but the quick reminder of where the case stands now is
- In 2024 the court said a vendor that performs core hiring functions for employers can be their “agent” and sued directly under Title VII, the ADEA and the ADA.
- In May 2025 the court preliminarily certified a nationwide ADEA collective. In February 2026 it authorized notice to everyone 40 and over who applied through Workday’s platform since September 24, 2020. The opt-in window closed March 7.
- On March 6, 2026, the court rejected Workday’s argument that ADEA disparate-impact protection only reaches current employees, not applicants.
- On June 22, 2026, the court let the California FEHA claims and a disability “proxy” claim go forward. The proxy theory is that screening on things like employment gaps is screening on medical leave and age. Think about what that does to every résumé filter in your ATS.
- Workday has asked for an interlocutory appeal of the certification ruling, so the scope could still change. The case has not gotten smaller yet.
But then the Court issued its May 29, 2026 discovery order, which every privacy lawyer should read twice. Magistrate Judge Beeler ruled two things. First, Workday’s bias-testing data stayed privileged, because lawyers curated it and used the results to give legal advice. Facts are still discoverable, but counsel’s selection and framing were not.
Second, the plaintiffs could not force Workday to produce its customers’ applicant data, because Workday’s subscription agreements gave it no legal right to demand that data from customers. No right to demand, no Rule 34 “control.” A confidentiality clause written for data protection worked as a discovery shield. I’ll come back to why that matters.
Interestingly enough, new cases are not pleading discrimination as a cause of action at all. Take Kistler v. Eightfold AI, filed in January 2026, for example. It is not a bias case. Rather, it is a Fair Credit Reporting Act case. The complaint says Eightfold scraped data on over a billion workers, scored applicants on a zero-to-five scale, and tossed the low scores before a human ever looked. The theory is that the score is a “consumer report,” which means disclosure, authorization, adverse-action notices and dispute rights, whether or not the model is fair. It was brought by a former EEOC chair. If that theory holds, the compliance architecture for AI screening becomes FCRA’s, and that is a lot more mechanical and a lot easier to prove than disparate impact.
Harper v. Sirius XM, Case No. 2:25-cv-12403 (E.D. Mich. August 4, 2025), is aimed at the employer, not the vendor. Roughly 150 rejections, and the allegation is the applicant tracking system leaned on ZIP code, school and work history as race proxies. Illinois has since banned ZIP-code proxies by name. The plaintiffs’ theory got written into a statute.
Baker v. CVS Health Corporation, Case NO. 2-11483(D.C. Mass. Feb. 16, 2024), alleged that an AI video-interview tool worked as a lie detector under a state statute written decades before anyone imagined the technology because part of its functionality was to determine if an applicant was being honest.
In the ACLU’s charge against HireVue and Intuit (EEOC, March 2025), the EEOC alleged that a deaf Indigenous applicant who was subject to an AI video interview with no captioning was discriminated against when she was a denied request for real-time captioning. The open question is whether your accommodation duties attach to the tool the same way they attach to a human interviewer. I do not think it is a longshot to predict that the answer is going to be “yes”.
And the one that is really a monitoring case is In re Otter.ai Privacy Litigation, Case No 5:25-cv-06911 (N.D. Cal. August 13, 2026), in which the Northern District of California let the federal Wiretap Act, California Invasion of Privacy Act, and Illinois biometric claims go forward against an AI meeting notetaker. The court tossed the computer-fraud theory, but the core survived. The argument is that a bot that joins a call, sends audio to a vendor’s servers, trains on it and builds speaker profiles is a third-party eavesdropper, not a passive tool the host controls. Cruz v. Fireflies.AI, Case No. 3:25-cv-03399 (C.D. Ill.) alleges the Fireflies.AI’s meeting assistant recorded, analyzed, transcribed and stored participants’ voices, including non-uses, without providing the notice, written consent, and retention safeguards required by Illinoi’s Biometric Information Privacy Act. Lastly, Basich v. Microsoft Corporation, Case No. 2:26-cv-00422 (W.D. Wash. Feb. 5, 2026), is the first major BIPA class action alleging that Microsoft Teams’ Live Transcription creates biometric “voiceprints” via speaker diarylation without Illinois-required notice or written consent.
This is an HR problem because that notetaker is sitting in your candidate interviews, your performance conversations, and your workplace investigations. Roughly a dozen states require all-party consent to record. A single interview panel can generate a dozen voiceprints. And here’s the trap: a spoken “this call is being recorded” at the top of the meeting may satisfy a wiretap statute but it does not satisfy BIPA, which wants written notice and a written release before a voiceprint is captured. Ask your recruiting team whether they let a notetaker into interviews. In most companies I talk to, nobody has checked.
California rewrote the monitoring rule on September 30, 2026
On September 30, 2026, California rewrote the monitoring rulebook when Governor Newsom signed four workplace AI bills. Recall last year when he vetoed SB 7, the original “No Robo Bosses Act,” saying it was too broad and the privacy agency’s automated-decision regulations would cover some of the same ground. The legislature took the hint and came back narrower and split the idea into separate bills. He signed all of them.
The useful way to think about these is not bill by bill. It is by where the monitoring happens. Because the tools do not respect the boundary between the application portal and the warehouse floor, or between the office and someone’s living room, and neither do the new laws.
The interesting thing about the four new bills is that only three of the four cover current employees. SB 947 expressly excludes applicants. AB 1883 dropped independent contractors in the final version and does not mention applicants. AB 1331 covers employees. So if you are looking for the rules on AI that screens, ranks and watches job candidates in California, they are in the layer that already exists:
- The Civil Rights Department’s FEHA regulations on automated decision systems, in effect since October 1, 2025. The “vendor did it” is not a defense. An agent acting for you is covered. You keep ADS records, including inputs, outputs, scores, rankings, selection criteria and bias-testing results, for four years. And whether you tested for bias, or did not, is evidence.
- The Privacy Protection Agency’s ADMT regulations, with full compliance due January 1, 2027, including for tools you are already using. Pre-use notice, opt-out rights, a right to meaningful information about the logic, and documented risk assessments signed by a senior executive. If you were already using the tool before 2026, the assessment is due by the end of 2027 and you attest to the agency by April 1, 2028.
And on top of that layer sit the cases above. The applicant side of the house is governed by civil rights regulations, privacy regulations, and a plaintiffs’ bar that has figured out FCRA, wiretap and biometric law work when discrimination is hard to plead.
While you are at work
SB 947 is the “No Robo Bosses Act” effective July 1, 2027. With it, California becomes the first state to require a human to sign off before an employer “primarily relies” on an automated decision system to discipline or fire someone. It requires three things:
- You cannot rely solely on the tool. And “tool” is broad: any computational process from machine learning, statistical modeling, data analytics or AI that spits out a score, classification or recommendation used to assist or replace a human’s judgment. Productivity scoring, attendance tracking, route optimization, quality monitoring, customer review aggregation, performance dashboards. Spam filters and spreadsheets are carved out. Not much else is.
- If you primarily rely on the output, a human must corroborate it using the data that produced it or other real evidence: supervisory evaluations, personnel files, work product, peer reviews, witness interviews. If the reviewer cannot corroborate it, or concludes it is inaccurate, incomplete or misleading, you cannot use it. At all.
- At the moment you deliver the decision, you hand the employee a standalone, plain-language written notice, in whatever language you routinely use with your workforce, that says you primarily relied on an automated system, that a human corroborated it, that retaliation is prohibited, and who they can call to have a person explain it. Separately, once a year an employee can ask for a meaningful description of the last 12 months of their own data the system used.
There are two provisions people keep skipping that I think demand more attention. First, the statute also prohibits using an ADS to infer an employee’s protected status, or to predict that a worker will exercise a legal right and take adverse action on that prediction. A model that flags “likely to file a complaint” or “likely to organize” is now something you cannot act on in California.
Second, there is a burden shift. Once it is shown that an automated system was used in a disciplinary decision, you have to prove either that you did not primarily rely on it or that you did the human review and gave the notice. If your reviewer’s file is a checkbox that says “agreed,” you will lose.
It is $500 per violation, enforced by the Labor Commissioner and public prosecutors, with injunctive relief, punitive damages and attorneys’ fees on the table. There is no express private right of action, but the burden-shift language refers to “any civil action,” and you can guess who’s going to test that. Unionized employers can bargain out of it with a clear and unambiguous waiver. Notably, the statute is clear that complying with SB 947’s notice satisfies substantially similar notice duties under other California laws, but it does not satisfy the privacy agency’s ADMT regulations.
AB 1883 effects the emotion and neural data ban. Effective January 1, 2027, a company cannot use a workplace surveillance tool that uses AI to recognize, infer or predict an employee’s emotional state, or to collect “neural data” from their nervous system. The definition of “workplace surveillance tool” is anything that collects employee data, activities, communications, actions, biometrics or behaviors by means other than a person watching. Video and audio, continuous incremental time-tracking, geolocation, electromagnetic and photo-optical tracking are listed, and the list is not exhaustive.
This prohibition includes sentiment and tone analysis in a contact center platform and the “engagement” or “burnout” score in a collaboration or HR analytics tool, both of which are standard in an HR tech stack these days. Further, anything that reads email, chat or call transcripts for mood is likely covered, because the definition says “communications.” On the neural side, it is EEG headbands, fatigue-monitoring caps, focus-tracking earbuds. The law excludes data inferred from non-neural sources, so a step count or a general heart-rate reading probably is not neural data. But a device that uses AI to infer stress from those readings runs into the emotion prohibition anyway.
There is a safety exception, and its scope is contested. Read narrowly, a safety tool stays legal only if it does not do the prohibited things. Read broadly, which the text arguably supports, a fatigue monitor on a crane operator might be fine even though it collects neural data. Until a regulator or a court tells us, document the safety purpose for anything you keep. Penalties are up to $500 per violation, and the statute does not say what one violation is. Per employee? Per day? The bill dropped its private right of action, but there is a line referring to a penalty “paid to the employee,” and PAGA lawyers read statutes for a living. One more thing: under the CCPA, neural data is sensitive personal information. And if you operate in Europe, the EU AI Act has banned emotion recognition in the workplace since February 2025, so you may already have a global configuration to borrow.
AB 1331, the bathroom bill. Effective January 1, 2027, you cannot use a workplace surveillance tool to monitor employees in a workplace bathroom unless a court orders it. Everyone’s first reaction is “of course not, who does that.” Then you read the definition, which is the same broad one as AB 1883, and realize it reaches your access badges, RFID scanners, Bluetooth beacons, WiFi positioning, mobile device management on company phones, warehouse wearables, and any time-and-attendance platform that logs idle intervals. The only categorical exclusions are smoke and carbon monoxide detectors and weapons screening.
Employees get the right to set the tool down before they walk in. Unclip the badge, leave the phone, take off the wearable. A company cannot discipline them for it. The exceptions are narrow: employees who have to stay reachable during breaks under existing law, a one-time badge swipe to get through a secured restroom door, and ID or safety devices your policy requires people to carry, but only if the device has no microphone, no camera and no embedded AI. A lot of modern badges and panic buttons fail that test. In addition, a company cannot aggregate the swipes to figure out how often or how long someone is in there. The earlier versions of the bill covered breakrooms, locker rooms, lactation spaces and off-duty hours. All of those were scrapped, but I would expect them back, one way or another.
This law has the same enforcement structure as AB 1883. Same $500. Same PAGA question. And the likeliest first claim is not some grand surveillance scheme, but rather, a manager on a power trip who pulls badge data to build a case that somebody takes too many breaks.
When you leave the building, and when you leave the job
This is where the “in and out of the workforce” framing comes in. The newest laws are about where monitoring follows people. Here are some of the most interesting ones I would expect to gain traction in other states:
- Home, car and personal phone: On January 11, 2026, Maine enacted “An Act to Regulate Employer Surveillance to Protect Workers.” It prohibits employers from using audio or video to monitor an employee in an employee’s home, personal vehicle or personal property unless the job actually requires it. An employee can refuse to install monitoring or data-collection apps on their personal device, which is going to make a lot of BYOD policies very interesting. Employers also have to tell applicants during the interview process that they monitor, and give every employee written notice at least once a year. Security cameras and GPS on company vehicles are carved out.
- The layoff: SB 951. Effective January 1, 2027 in California, if a mass layoff, relocation or termination is caused “in whole or in substantial part” by AI or other automation, the Cal/WARN notice must have four new elements. The words “This notice is for a technology displacement” must be at the top. The number, classification and location of the affected positions but be included, along with the job functions being automated. And the category of technology responsible. For those of us who have to draft WARN notices, this is a quite a new task.
Think about what that document is. It is a written, government-filed, publicly summarized statement of causation, signed by the company, saying the machine took the jobs. Unions, reporters, and plaintiffs’ lawyers alike will all be reading these notice.
Connecticut already has a lighter version in force as of October 1, 2026, which requires any WARN notice expressly say whether the layoffs relate to AI or technological change.
And one piece of good news: SB 690
I defend California Invasion of Privacy Act (CIPA) cases, so I will say this plainly. SB 690 kills private lawsuits over website and app “pen register” and “trap-and-trace” claims starting January 1, 2027, leaves enforcement to the Attorney General, and applies retroactively to pending claims filed within the prior two years. That is real relief for anyone who’s gotten one of the thousands of demand letters over ordinary cookies and pixels. It is not the end of CIPA. It does not touch the Section 631 wiretapping claims that drive most of the current litigation, and the retroactivity provision is going to be challenged. If you have got a pending claim, this changes your leverage, not your exposure.
It is not only California
California gets the headlines, but the monitoring and AI rules are stacking up everywhere, and they do not line up with each other.
- Connecticut has two laws hitting at once. The electronic monitoring statute that has been on the books since 1998 was rewritten by Public Act 26-73, which became effective on October 1, 2026. Now, notice has to say not just what kinds of monitoring you do but where, and it has to be posted in the place where the monitoring happens. New hires get a plain-language statement of what is prohibited and may be monitored without further notice. Separately, the Connecticut AI Responsibility and Transparency Act (Public Act 26-15) brings in an automated employment decision framework, amends the state fair employment law to say using an AI tool is not a defense to a discrimination claim (though documented testing can mitigate), and adds the WARN AI disclosure, all as of October 1, 2026. Pre-decision notice for new systems starts October 1, 2027.
- New York and Delaware have had monitoring-notice laws for years. New York wants notice at hire, a signed acknowledgment, and a posting. Delaware takes a one-time acknowledged notice or a daily login reminder. If you have people in those states and rolled out an AI tool that scores productivity or transcribes calls in the last two years without updating your notice, that is the top of your audit list.
- Illinois has been live since January 1, 2026 with HB 3773, which makes it a civil rights violation to use AI that has the effect of discriminating in employment decisions, to use ZIP code as a proxy, or to fail to tell people AI is in use. The implementing rules still aren’t final, which means employers are complying with a live obligation whose notice mechanics are still being written. The 2019 video interview law is still in force too, all with the BIPA sitting there as well.
- Colorado’s replacement law, SB 26-189, takes effect January 1, 2027. It regulates automated decision-making technology that “materially influences” a consequential decision, and consequential expressly includes compensation and scheduling. That pulls merit-pay models and shift algorithms into scope. The exemptions for federally regulated entities were removed, so banks, insurers and health plans that thought Colorado did not reach them need to re-run that analysis.
- New Jersey’s Division on Civil Rights adopted disparate-impact regulations on December 15, 2025 that specifically call out résumé screeners, scheduling filters and video analysis. The provision to read twice: reasonable steps are not a defense if they do not actually prevent the disparate impact. So, in Connecticut, documented testing reduces your exposure. In New Jersey, it does not excuse the outcome. Same program, two legal effects.
- New York City’s bias-audit law has been in effect since 2023, and in December 2025 the State Comptroller audited the city’s enforcement and called it “ineffective.” Three-quarters of test calls to 311 never reached the enforcing agency. Reviewing 32 posted audits, the agency found one problem; the Comptroller found seventeen. The agency has now promised proactive enforcement. If you have been quietly out of compliance for three years and nothing happened, that wasn’t a compliance record. It was an enforcement gap, and it is closing.
- Europe moved its dates, not its obligations. The Digital Omnibus pushed the AI Act’s high-risk employment rules to December 2, 2027. But the ban on emotion recognition in the workplace has been in force since February 2025, and the transparency obligations landed August 2, 2026 as scheduled.
- On the watch list: Michigan’s HB 5579, introduced this session, would add monitoring limits and AI transparency requirements. It has not moved yet, but given how the Maine and California bills traveled, I would not bet against it.
Privacy and employment law stopped being separate programs
For a privacy audience, this is the real headline of 2026, and here is why:
- California’s privacy agency, not its labor agency, wrote the rules on pre-use notice, opt-out and risk assessments for AI that makes significant employment decisions. The Governor cited those rules as a reason to veto last year’s labor bill. The privacy regulator is now doing employment law, and the labor statute signed this year says in so many words that complying with it does not get you out of the privacy rules.
- The CPPA’s ADMT regulations are also triggered by using automated processing to infer things like aptitude, performance, reliability, location or movements from systematic observation of employees. That is not a hiring rule. That is a monitoring rule, and it has been sitting in a privacy regulation the whole time.
- AB 1883 makes neural data a labor-law problem; the CCPA already made it sensitive personal information with notice-at-collection, right-to-limit and risk-assessment consequences.
- California’s civil rights regs say keep four years of ADS records including bias-testing results. Colorado says three. The CPPA says keep risk assessments five. Your privacy program’s data-minimization and deletion schedule will walk you straight into an EEO records violation if nobody reconciles them.
- And Mobley says whether anyone ever sees your bias-testing data depends on who directed the work and what your vendor contract says about demanding data. Those are privacy and procurement decisions that have become litigation strategy.
Put the last four together and you get the bind I keep walking clients through. Connecticut rewards you for testing. New Jersey holds you liable for the outcome whether you tested or not. California makes you keep the results for four years. And Mobley says the plaintiffs’ bar sees them, or does not, based on contract language and privilege decisions you made before the test ever ran. Those aren’t four separate questions. They are one design decision, and it has to be made at the beginning.
What I would be doing about it
None of this requires a 50-state program built from scratch. The differences between these laws are mostly timing, notice wording and who enforces. The underlying controls are the same. Here is where I would focus:
- Build the monitoring map. Not a tool inventory, a map. Every system that collects data about a person or produces a score, rank or recommendation about them, and where it follows them: application, interview, badge-in, desktop, phone, vehicle, home, exit. Include the notetaker. Include the thing HR bought on a credit card that legal never saw.
- Turn off what you cannot defend. Sentiment, mood, engagement, stress and burnout features for California employees, or everywhere if that is simpler. Restroom dwell-time and break-frequency reports. Neural wearables unless you have written down a specific safety purpose. Then confirm with the vendor, in writing, that it is actually off. The California bills focus on whether a tool is used for the prohibited purpose, not what it is capable of, so a documented disable is a real compliance step.
- Ask every vendor three questions in writing. Does your product infer or predict anyone’s emotional state from text, voice or video? Does it collect data measuring nervous-system activity? Does it produce a score, classification or recommendation about an individual? Then get the answers, plus notice before new features turn on, a right to disable them, testing cooperation, and model-change notice, into the contract before renewal.
- Decide what “primarily relies” means in your shop and write it down. SB 947 does not define it. Pick which tools you will treat as triggering the law and which are background, document why, and apply it consistently. Then build the human-review file: who reviews, what evidence they must look at, what they concluded about the system’s accuracy, and when. The burden shift means the file is your only defense.
- Draft the notices now, as one architecture with state switches. The SB 947 post-use notice (standalone, plain language, translated, human contact identified by role so it survives turnover). The CPPA pre-use notice. Connecticut’s location-specific posting. New York’s and Delaware’s acknowledgments. Maine’s annual notice and the interview disclosure. They are the same thing, they are just packaged differently.
- Write the WARN playbook before you need it. Who decides whether a reduction is “in whole or in substantial part” caused by AI? What’s the evidence? Who signs? Remember that in California it goes to a public quarterly report and in Connecticut it goes on the notice. Get HR, legal, ops and communications aligned on the answer before the first notice is drafted, because you won’t have time after.
- Fix retention. Reconcile the EEO rules (keep four years in California, three in Colorado) with the privacy program’s deletion schedule. When they conflict, keep.
- Make the privilege call before the test runs. Decide who directs bias testing, under whose privilege, and what the vendor agreement says about who can demand the data. Most companies will end up with two tracks: a published, audit-grade assessment that satisfies the disclosure laws, and a separate privileged workstream for exploratory analysis and remediation. The one thing you cannot do is figure this out after the complaint arrives.
- Do not calibrate to Texas. Texas’s AI law only prohibits intentional discrimination and says disparate impact alone does not prove intent. That is the most permissive rule in the country, and it does not change Title VII, the ADA, the ADEA or the law of the other 49 states. A national employer that builds to Texas is not going to be protected anywhere else.
- Train the managers. The badge data is off limits for break discipline. The sentiment dashboard is off limits for performance reviews. A flag from the system is a reason to look harder, not a decision. And the notetaker stays out of the interview until somebody has read the consent law for every state on the call.
Takeaways
Last year the argument was that the states had stepped into a void. This year the point is sharper. The states did not just fill the hiring gap. They drew the first real lines around watching people at work: no reading their emotions, no tracking them into the bathroom, no filming them in their own kitchen, no firing them on a machine’s say-so without a person checking the work, and no pretending the layoff wasn’t about automation. The federal government is arguing about those lines in court. It has not erased one. And the line that moved the most in 2026 is not about who gets the job. It is about what happens to you once you have got it.
About the Author: Sara Jodka, CIPP-US, CIPP-E, and CIPM (Member, Columbus) is a member of the firm’s labor and employment department where she combines her privacy and technology expertise to workplace matters, including employee privacy. She can be reached at 614-744-2943 or SJodka@dickinsonwright.com.