New Jersey's S4075 would regulate AI workplace surveillance and quietly reshape proof in every workers ' compensation claim.
The most protective workplace artificial intelligence bill in the country is pending in the NJ Legislature, and the workers' compensation bar should be participating in the discussion.
Senate Bill 4075, introduced May 4, 2026 by Senators Andrew Zwicker and Shirley K. Turner and now carrying eight Senate sponsors, would regulate three categories of technology across every New Jersey workplace: the electronic monitoring tool, or EMT; the automated employment decision system, or AEDS; and the automated benefit or service decision system, or ABSDS. The Assembly companion, A4981, was introduced May 7, 2026, with eleven sponsors. Both bills remain in committee. Neither has moved since spring.
Read as an employment bill, S4075 is a privacy and collective bargaining measure. Read as a workers’ compensation lawyer reads it, it is something else entirely. It is a statute about who gets to collect the biological record of a working body, who gets to keep it, how long they get to keep it, and what happens to that record when the worker is hurt and files a claim petition. That is an evidence statute wearing a labor statute's clothing.
What the Bill Actually Does
The architecture is more extensive than anything enacted in the United States to date.
• Pre-deployment gate. No AEDS or EMT may operate until an independent auditor, or the Department of Labor and Workforce Development where public employees are involved, completes an impact assessment confirming compliance, describing training-data disparities, and analyzing accuracy, reliability, validity, and error rates. Section 3.
• Public registry. The full assessment and an accessible summary go to the Department within sixty days for a public registry available to affected employees, applicants, and their authorized representatives. Section 3(g).
• Purpose limitation. Monitoring is confined to six allowable purposes, must use the least invasive means, must cover the smallest number of employees and the least data needed, and must sample no more often than necessary. Section 3(b).
• Hard prohibitions. No monitoring in bathrooms, break rooms, sick rooms, wellness rooms, or lactation areas. No surveillance while off duty, on leave, or at meal or rest breaks. No surveillance of a residence or personal vehicle. No forced installation of monitoring software on personal devices, and an absolute right to refuse without retaliation. No implanted or subcutaneous devices. No adverse action based solely on keystroke logging, idle-time trackers, or mouse-movement monitors. Section 2(c) through 2(i).
• Biometric, health, and wellness data. These may not be transferred to any third party or government entity absent a legal requirement, may not be used in any employment-related decision, and may not be retained after employment ends. Section 2(k).
• Human oversight. No decision affecting terms or conditions of employment may rest exclusively or determinatively on system outputs. Trained internal reviewers must corroborate the data and must have authority to reject it. Section 9.
• Notice and appeal. Sixty days' notice before implementation, ten days' notice before an adverse decision takes effect, and a right to review and copy every datum used, plus a clear explanation of how the system produced the output, including the weighting of factors. Sections 6 through 8.
• Enforcement. Liquidated damages up to 200 percent of lost wages, civil fines, punitive damages, fees, a disorderly persons offense for knowing and willful violations, joint and several liability running to the vendor, and a rebuttable presumption of retaliation for adverse action within ninety days of a complaint. Sections 18 through 20.
The act would take effect eighteen months after enactment. Systems already in service on the effective date get six months to complete an impact assessment, twelve months in the case of an ABSDS.
Where New Jersey Stands Against California and the Rest
California moved first and moved narrowly. Assembly Bill 1883, by Assembly Member Isaac Bryan, was signed on September 3, 2026 and takes effect January 1, 2027. It adds Part 5.8 to Division 2 of the Labor Code and does two things: it bars an employer from using an AI workplace surveillance tool to recognize or infer an employee's emotional state, and it bars the collection of neural data, defined as information generated by measuring the activity of an employee's central or peripheral nervous system that is not inferred from nonneural information. Enforcement runs through the Labor Commissioner or a public prosecutor, with a civil penalty capped at $500 per violation. Two companions remain on the Governor's desk with a September 30 deadline: SB 947, which would prohibit sole reliance on automated decision systems for discipline or termination, and SB 951, which would extend the California WARN Act to AI-driven mass layoffs. A broader predecessor, the No Robo Bosses Act, was vetoed in 2025.
Illinois took the civil rights route. House Bill 3773, Public Act 103-0804, amended the Illinois Human Rights Act effective January 1, 2026 to make discriminatory AI use in employment a civil rights violation, to bar zip codes as protected-class proxies, and to require notice. It mandates no audit and no impact assessment, and the Department of Human Rights withdrew its proposed notice rules on June 2, 2026, leaving the statutory duty in force without implementing detail.
Colorado retreated. Senate Bill 24-205, the first comprehensive state AI act, imposed a duty of care and mandatory algorithmic impact assessments. Its effective date was pushed from February 1 to June 30, 2026 after a failed special session. Then a federal magistrate stayed enforcement in April 2026 amid a constitutional challenge, and on May 14, 2026 Governor Polis signed SB 26-189, which repealed and replaced the act outright with a disclosure-and-rights framework for automated decision-making technology effective January 1, 2027. The duty of care is gone. The bias audits are gone.
Washington, federally, is pushing the other way. Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence, issued December 11, 2025, directs the Attorney General to stand up an AI Litigation Task Force whose sole responsibility is to challenge state AI laws on preemption, dormant commerce clause, and First Amendment grounds, and directs Commerce to publish a list of state laws suitable for referral. A New Jersey statute with mandatory pre-deployment audits, a public registry, and joint and several vendor liability is exactly the profile that list was built to capture.
New Jersey already has a foothold that the other states lacked. In January 2025, Attorney General Platkin and the Division on Civil Rights issued Guidance on Algorithmic Discrimination and the New Jersey Law Against Discrimination, confirming that the LAD reaches algorithmic disparate treatment and disparate impact and that an employer cannot shift liability to its AI vendor. S4075 would supply the machinery the guidance presumes but cannot create.
The Biological Data Gap—Neural Data
Here the bill needs work, and the workers' compensation bar is the constituency best positioned to advocate.
S4075's definition of biometric information is broad on its face. It captures fingerprints, voice prints, retinal and iris scans, gait, facial maps and geometries, facial recognition, voice analysis, emotion recognition technology, genetic information, and a catch-all for other unique biological, physical, or behavioral patterns or characteristics. Gait and emotion recognition are meaningful inclusions. A gait signature is a musculoskeletal record. Emotion recognition is a proxy for psychological state.
But the definition never uses the words neural data. That is not a cosmetic omission. Colorado's HB 24-1058 defined neural data expressly in 2024. California added it to the Consumer Privacy Act through SB 1223, then wrote it into AB 1883 as a stand-alone employment prohibition. Montana amended its genetic privacy statute to require a warrant for law enforcement access. Connecticut folded it into its data privacy act. The patchwork is real, and New Jersey would be legislating into it without naming the category.
Worse, the definition carries an express carve-out: biometric information does not include photographs, audio or video recordings, or data or information generated from photographs or audio or video recordings. Read literally, that exclusion swallows a growing share of biological monitoring. Remote photoplethysmography extracts heart rate and heart rate variability from ordinary video of a face. Computer-vision fatigue detection scores blink rate, eyelid closure, and head position from a cab-facing camera. Pose-estimation systems derive lifting mechanics, trunk flexion, and repetitive-motion counts from warehouse video. Each of those is a physiological measurement. Each of them is, on the statute's own words, data generated from a video recording, and therefore outside the biometric protections of Section 2(k).
That means the most consequential biological surveillance in modern industrial workplaces, the surveillance that generates the exact physiological record an injured worker will later need, falls through the definitional floor.
Four Ways This Reaches Compensation Claims
1. Quotas, causation, and the intentional wrong exception
Section 2(h) prohibits deploying an EMT or AEDS in a manner that harms or is likely to harm worker health and safety by setting, or facilitating the setting of, productivity quotas or performance standards likely to contribute significantly to harming worker health and safety. Section 6(a)(5) requires that every quota be disclosed in advance, and Section 6(d) bars any adverse decision based on an undisclosed quota.
For the cumulative-trauma claim, this is causation evidence handed over by statute. An algorithmically set rate, disclosed in writing, correlated against injury incidence, is precisely the objective proof of exposure that Fiore v. Consolidated Freightways, 140 N.J. 452 (1995), demands when it requires that the disease be due in a material degree to conditions characteristic of the employment. It also raises the harder question. Where an employer sets a machine-driven pace, receives the impact assessment warning that the pace is likely to harm health and safety, and runs the pace anyway, the analysis under Millison v. E.I. du Pont de Nemours & Co., 101 N.J. 161 (1985), and Laidlow v. Hariton Machinery Co., 170 N.J. 602 (2002), begins to look different. A documented, pre-deployment, written warning of substantially certain harm is not the ordinary industrial risk the exclusivity bargain contemplates. Van Dunk v. Reckson Associates Realty Corp., 210 N.J. 449 (2012), still demands a virtual certainty, and most claims will not clear it. But the registry created by Section 3(g) would, be an evidential link.
2. Biometric and wellness data in return-to-work decisions
Section 2(k) prohibits using biometric, health, or wellness data or information in making an employment-related decision, and prohibits retaining it after employment ends. In practice that reaches the wearable that logs an employee's heart rate, the fatigue monitor in the truck cab, and the exoskeleton telemetry that scores lifting form. An employer could still collect that data for a Section 3(a)(5) safety purpose. It could not use the data to decide whether the returning claimant gets the light-duty slot.
The parallel to Millison is exact. There, the employer's own medical department discovered asbestos-related disease on annual films and said nothing, and the Court held that the fraudulent concealment of already-discovered disease fell outside the exclusivity bar. A modern employer's monitoring stack is a medical department that never closes. It sees the degradation before the worker does. Section 2(k) forbids the employer from acting on it. It fails to require that the employer reveal the data to the employee.
3. The thirty-seven-month destruction clock
Section 5(b) requires three years of record retention and then commands destruction of the data no later than thirty-seven months after collection, absent uncoerced written consent to retain.
Set that clock against New Jersey's occupational disease limitations period. Under N.J.S.A. 34:15-34, there is no time limitation on filing an occupational disease claim, subject to a two-year period running from the date the claimant first knew the nature of the disability and its relation to the employment. Earl v. Johnson & Johnson, 158 N.J. 155 (1999), makes plain how far into the future that trigger can sit. Fiore and the cumulative-trauma line assume the claimant can eventually reconstruct exposure.
A thirty-seven-month destruction mandate destroys that reconstruction. The ergonomic telemetry, the quota logs, the air-monitoring feeds, and the validation records that would prove a twenty-year repetitive-strain or inhalation exposure are gone before the disease is diagnosed. The bill's privacy instinct is sound. Its interaction with occupational disease law is not.
There is a second-order problem. Once a claim petition is filed, the common-law duty to preserve attaches. Rosenblit v. Zimmerman, 166 N.J. 391 (2001), supplies the adverse inference, discovery sanctions, and fraudulent concealment remedies for destruction of relevant evidence. An employer that destroys monitoring data on a statutory schedule after notice of a claim will be arguing statutory command against preservation duty, and will not enjoy that argument.
4. The algorithm that learns to spot an injured worker
This is the risk that has already materialized in federal court. In Mobley v. Workday, Inc., No. 23-cv-00770-RFL (N.D. Cal.), Judge Rita Lin's order of June 22, 2026 sustained an Americans with Disabilities Act claim built on the theory that algorithmic hiring tools identify and rely on proxy indicators of illness or health-related limitations, including medical-related leave and patterns consistent with treatment and recovery, and screen applicants out on inferred health status rather than job-related qualifications. The court had earlier held, on July 12, 2024, that the vendor could be an employer under an agency theory.
The injured worker is returning to the labor market after an injury, and the employer may try to us use an algorithm. Section 2(j) of S4075 would prohibit an employer or vendor from using an AEDS to obtain, infer, analyze, or use protected classifications or characteristics not directly related to work performance or qualifications, and Section 2(k) would bar health data from employment decisions. New Jersey's DCR guidance already establishes that the LAD reaches the outcome. S4075 would supply the audit trail needed to prove it.
Evidentiary Consequences in the Division
Compensation practice is unusually exposed here because the Division's evidentiary posture is unusually permissive. Under N.J.S.A. 34:15-56, the judge of compensation is not bound by the Rules of Evidence, though findings must rest on competent evidence. Machine-generated output, algorithmic scores, and vendor-supplied risk ratings will therefore arrive in the Division with a lower barrier than they would face in the Law Division.
Three consequences follow.
1. Authentication becomes the fight. An AEDS score is not a business record in any ordinary sense. It is a model output whose meaning depends on training data, thresholds, and the weighting of factors. Section 8(b)(1) of S4075 would give the worker the right to review and copy every datum used, the impact assessments, the oversight record including whether a human modified the output and how, and a clear, complete explanation of how the system produced the output, including the weighting of factors. The legislation is a discovery template. Petitioner's counsel should be using its language in demands whether or not the bill passes.
2. Error rates become cross-examination. Section 3(e)(5) requires the vendor to supply an analysis of accuracy, reliability, validity, and error rates, including the foreseeable effects of tuning and retraining, and Section 3(g) puts the report in a public registry. When a respondent's expert leans on an algorithmic functional-capacity score or an AI-generated utilization review, the published error rate is available to impeach it. An expert who relies on an output without the validation record behind it is offering exactly what Townsend v. Pierre, 221 N.J. 36 (2015), calls a net opinion: a conclusion without an identified factual basis or a demonstrated methodology.
3. Geolocation cuts both ways. Location telemetry is often the cleanest proof of whether an injury arose out of and in the course of employment, the question Hersh v. County of Morris, 217 N.J. 236 (2014), resolves by asking about employer control of the worksite. Section 2(d) would switch that telemetry off outside working time, protecting privacy while also deleting the proof that a parking lot fall or a mobile worker's roadside injury occurred where the claimant alleges.
One more structural point. McDonald v. Symphony Bronzeville Park, LLC, 2022 IL 126511, held that biometric privacy claims for statutory damages are not the kind of injury that fits within the compensation act, so exclusivity did not bar them. S4075's Section 19 civil action, its liquidated damages, and its punitive damages create the same structure in New Jersey: a privacy claim running alongside, not through, the Division. Practitioners should expect parallel tracks, not a single forum.
What S4075 Does Not Reach
The largest gap is the one nearest to this practice. S4075 regulates employers, public entities, and their vendors. It regulates automated decisions about public benefits and services. It does not regulate the carrier or third-party administrator that uses artificial intelligence to adjudicate a workers' compensation claim, to set a reserve, to score a file for settlement, or to run a utilization review of requested treatment.
That omission is not unique to New Jersey. The wave of state statutes requiring physician review of AI-generated adverse determinations- California's SB 1120, Texas SB 815, Illinois HB 2472, and their counterparts in Arizona, Maryland, Nebraska, Indiana, Utah, and Washington- generally exclude workers' compensation from their scope along with ERISA self-insured plans. In most of the country, an injured worker can have an algorithm deny surgical authorization with no statute standing in the way.
An injured worker in New Jersey would, under S4075 as drafted, gain a right to a human reviewer for a hiring decision and keep no such right for the denial of a lumbar fusion.
Amendments Worth Making
• Name neural data. Add an express definition of neural data, information generated by measuring the activity of the central or peripheral nervous system, to the biometric information definition, on the Colorado and California model.
• Close the video carve-out. Limit the photograph and video exclusion to the recordings themselves, and bring physiological measurements derived from them, heart rate, fatigue scoring, pose and gait estimation, back inside the definition.
• Toll the destruction clock. Suspend the thirty-seven-month destruction mandate on notice of a claim petition, an occupational disease exposure, or a written preservation demand, and extend the baseline retention period for exposure data to match the occupational disease limitations structure of N.J.S.A. 34:15-34.
• Reach the claim file. Extend the Section 9 human-oversight requirement and the Section 8 notice-and-appeal rights to automated adjudication of workers' compensation claims and to AI-assisted utilization review, whether performed by a carrier, a third-party administrator, or a self-insured employer.
• Add an affirmative disclosure duty. Where an employer's monitoring generates a physiological finding suggesting injury or disease, require that the finding be disclosed to the worker. Section 5(a)(2) requires accuracy and access on request. Millison teaches that access on request is not enough when the employer knows and the worker does not.
What to Do Now
The practice consequences do not wait for legislative enactment. Some actions should be taken now
• Serve monitoring-data demands at intake. Wearables, telematics, badge and access logs, productivity dashboards, quota records, and camera-derived analytics. Ask for retention schedules and destruction dates in the same demand.
• Send a litigation hold letter the day the petition is filed, naming the systems by category. Rosenblit is the leverage.
• In any case involving a productivity standard, demand the quota, the method by which it was set, and any vendor validation or risk documentation supporting it.
• Where the respondent offers an algorithmic score, a functional-capacity model, or an AI-assisted utilization review, demand the error rate and the validation record, and treat the absence of one as a Townsend problem.
• For claimants re-entering the labor market, preserve the rejection pattern. Mobley makes proxy-inference screening a litigable theory, and the LAD reaches it in New Jersey under the DCR guidance.
• Comment to the Senate Labor Committee on S4075 and the Assembly Science, Innovation and Technology Committee on A4981. The workers’ compensation bar has data nobody else in that room has.
The bargain struck in 1911 was that the worker surrendered the tort remedy in exchange for swift and certain compensation. That bargain assumed the worker could prove what happened to his body at work. For a century the proof was testimony, a supervisor's memory, and a treating physician's chart. It is now a data exhaust that the employer owns, the vendor controls, and a statute may soon require be deleted at month thirty-seven.
Regulate the harvest by all means. Just do not delete the evidence.
Sources
1. New Jersey Senate Bill No. 4075, 222nd Leg. (introduced May 4, 2026), Zwicker and Turner, sponsorship updated June 15, 2026. Bill page and status; full text.
2. New Jersey Assembly Bill No. 4981, 222nd Leg. (introduced May 7, 2026), companion to S4075. A4981.
3. "Implementing AI While Regulations Are Pending," New Jersey Business Magazine, June 2026 (Senator Zwicker on S4075). NJBIZ.
4. California Assembly Bill 1883 (Bryan), Workplace surveillance tools, adding Part 5.8 (commencing with § 1580) to Division 2 of the Labor Code, signed September 3, 2026, effective January 1, 2027. Bill text.
5. Grant Larson, "Legislative Scoop: Key Employment Law Updates from California about Artificial Intelligence," GovDocs, Sept. 8, 2026 (AB 1883 signing date; SB 947 and SB 951 status). GovDocs.
6. "California Passes Bill to Ban AI Tracking of Worker Neural Data," Bloomberg Law, Aug. 31, 2026. Bloomberg Law.
7. California Senate Bill 947 (automated decision systems in discipline and termination). SB 947. California Senate Bill 951 (WARN notice for AI-driven layoffs). SB 951.
8. California Senate Bill 7 (2025), the No Robo Bosses Act, vetoed. Senate District 5 release.
9. Illinois House Bill 3773, Public Act 103-0804, amending the Illinois Human Rights Act effective Jan. 1, 2026; IDHR proposed rules withdrawn June 2, 2026. Seyfarth Shaw LLP.
10. Colorado Senate Bill 26-189, Automated Decision-Making Technology, signed May 14, 2026, effective Jan. 1, 2027, repealing and reenacting SB 24-205. Colorado General Assembly; Seyfarth Shaw LLP analysis.
11. Colorado Senate Bill 25B-004, extending the SB 24-205 effective date to June 30, 2026. Colorado General Assembly.
12. Colorado House Bill 24-1058, Protecting the Privacy of Individuals' Biological Data, signed act (neural data definition). Colorado General Assembly.
13. "Your Brain, Their Rules: The Growing Patchwork of Neural Data Regulation," Cooley LLP, Feb. 23, 2026. Cooley LLP.
14. Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence, 90 Fed. Reg. (Dec. 16, 2025). Federal Register.
15. New Jersey Office of the Attorney General and Division on Civil Rights, Guidance on Algorithmic Discrimination and the New Jersey Law Against Discrimination (Jan. 2025). Guidance; announcement.
16. New Jersey Department of Labor and Workforce Development, Division of Workers' Compensation seminar materials, Learned Treatises and Other Medical Proofs (evidentiary standard under N.J.S.A. 34:15-56). NJDOL.
17. "AI Healthcare Regulations: Federal, State and HIPAA" (state utilization-review statutes and their workers' compensation exclusions). Live Compliance.
18. Millison v. E.I. du Pont de Nemours & Co., 101 N.J. 161, 501 A.2d 505 (1985).
19. Laidlow v. Hariton Machinery Co., Inc., 170 N.J. 602, 790 A.2d 884 (2002).
20. Van Dunk v. Reckson Associates Realty Corp., 210 N.J. 449, 45 A.3d 965 (2012).
21. Fiore v. Consolidated Freightways, 140 N.J. 452, 659 A.2d 436 (1995).
22. Earl v. Johnson & Johnson, 158 N.J. 155, 728 A.2d 820 (1999).
23. Hersh v. County of Morris, 217 N.J. 236, 86 A.3d 101 (2014).
24. Rosenblit v. Zimmerman, 166 N.J. 391, 766 A.2d 749 (2001).
25. Townsend v. Pierre, 221 N.J. 36, 110 A.3d 52 (2015).
26. McDonald v. Symphony Bronzeville Park, LLC, 2022 IL 126511 (Feb. 3, 2022).
27. Mobley v. Workday, Inc., No. 23-cv-00770-RFL, Order Granting in Part and Denying in Part Motion to Dismiss (N.D. Cal. June 22, 2026), ECF No. 360.
28. Mobley v. Workday, Inc., No. 23-cv-00770-RFL, Order on Motion to Dismiss First Amended Complaint (N.D. Cal. July 12, 2024) (agency theory).
Recommended Citation
Gelman, Jon L., Your Body Becomes Evidence, WORKERS' COMPENSATION, workers-compensation.blogspot.com (Sept. 13, 2026), https://workers-compensation.blogspot.com/2026/09/your-body-becomes-evidence.htmll.
Jon L. Gelman of Wayne, NJ, is the author of NJ Workers' Compensation Law (West-Thomson-Reuters) and co-author of the national treatise Modern Workers' Compensation Law (West-Thomson-Reuters).
Blog: Workers' Compensation
LinkedIn: JonGelman
LinkedIn Group: Injured Workers Law & Advocacy Group
Author: "Workers' Compensation Law" West-Thomson-Reuters
Blue Sky: jongelman@bsky.social
Substack: https://jongelman.substack.com/
© 2026 Jon L Gelman. All rights reserved.
Attorney Advertising
Prior results do not guarantee a similar outcome.
No comments:
Post a Comment