AI is Not the Boss of Me: A Meaningful Human-Oversight Requirement for Workplace Decisions

Amazon’s software recommended terminations for warehouse workers whose productivity scores fell below algorithmically tracked thresholds.[1] Delivery drivers for the company also report being fired by an algorithm.[2] Those workers received termination notices without explanation and could not tell whether any human being had been involved in the process.

Examples like these from Amazon are not outliers. Rather, this approach reflects the current direction of American workplace management. Software that screens, ranks, disciplines, and terminates workers is commonly referred to as automated decision systems (“ADS”) and employers now deploy them at scale in warehouses, call centers, delivery platforms, hotels, and corporate offices across every state. Around 4 in 10 U.S. workers report that they receive their schedules or assignments from automated systems.[3] Workers subject to ADSs often do not know they exist; instead, they receive a warning or termination notice, but have no way to know whether a human being was involved in the decision, what data was used, or whether that data was accurate.

Introduction

Existing law has no mechanism for addressing potential unfairness in these decisions. Most workers are “at will” and so can be fired for almost any reason or no reason at all. Anti-discrimination law generally requires a worker to be able to show that discrimination occurred and to establish a causal link between the bias and the decision— a burden that ordinarily depends on access to information about how a decision was made.[4] Yet algorithmic management makes it substantially harder to prove these elements.

The principle of meaningful human oversight — sometimes described as keeping a human in the loop — for consequential employment decisions addresses this gap. Its core requirement is narrow: when an employer relies on an automated system to discipline or terminate a worker, a human reviewer with the authority and information to reach an independent judgment must be part of the process, and not merely a rubber stamp for the system’s output. Implementing this principle requires providing advance notice when such systems are deployed, giving workers the right to access their own data used in such decisions, and creating a structured appeal process. What this principle requires of employers is comparable to processes they already handle under existing laws and policies.

The political moment is strong. A 2026 AFL-CIO survey found that 95% of working registered voters support requiring a human to be the final decision-maker on matters affecting an individual worker’s employment.[5] State legislatures are already codifying the principle of meaningful human oversight in domains other than employment. And regulation of ADSs, including its use in the workplace, has already passed in states like Colorado, Illinois, and Connecticut. Those who lead on this issue will be on the right side of both public opinion and the arc of pro-worker AI regulation.[6]

1. The Problem: Algorithmic Management in Today’s Workplace

The Systems and What They Do

ADSs in the employment context are software that analyze workplace data to produce outputs that employers can use in personnel decisions, and their deployment has been rapid and largely invisible to the workers affected by them. ADSs perform three primary functions.

Ranking systems score or compare workers using productivity, manager evaluations, and customer ratings, generating alerts and recommendations that determine compensation, preferred shifts, or discipline flags. For example, Centrical collects real-time performance data to produce gamified worker rankings displayed on manager dashboards.[7] WorkTime monitors computer activity and generates distraction scores.[8] Harver uses behavioral assessments to rank job candidates.[9] Like the predictive and real-time systems described below, ranking tools are concerning because the scores that determine pay, shifts, and discipline are generated by opaque formulas the worker never sees, cannot verify, and cannot meaningfully contest — and because they can entrench existing bias while presenting it as an objective metric.

Predictive systems use personal information like location data, purchase history, survey responses, and HR records to forecast performance, stress tolerance, or likelihood of organizing activity. One company, Perceptyx, analyzes employee surveys and HR data to predict engagement, retention risk, and union vulnerability, which highlights to managers the risk of organizing activity.[10] Other tools assess applicants’ financial records as proxies for fraud risk.[11] Though there is nothing wrong with evaluating employee engagement, “accurately predicting people’s social behavior is not a solvable technology problem.”[12] And employers are deploying these tools at scale with no obligation to disclose their use.

Real-time direction systems use surveillance data to deliver task assignments and set pace with no human intermediary. In warehouses, algorithmic systems track worker movement down to the second, triggering automatic discipline when workers fall short of opaque productivity quotas.[13] Hotel platforms assign housekeeping tasks based on real-time check-in data, even if it requires workers to change floors constantly, and security platforms dispatch patrols algorithmically.[14] When the boss is an algorithm, there is no one to explain a decision, respond to a concern, or exercise judgment when something unusual occurs.

Who Bears the Burden of Algorithmic Management

The burden of algorithmic management is not felt equally. Sixty-five percent of White workers report at least some form of electronic monitoring at work, compared to 82% of Black workers and 73% of Hispanic workers.[15] Black and Hispanic workers are concentrated in the sectors where algorithmic management is most intensive and its consequences most immediate, such as warehousing, delivery, and food service.[16]

Many workers subject to these systems do not know they exist. They receive a rejection or a termination notice with no way to identify whether a human was involved, what data was used, or whether that data was accurate.

Why Existing Laws Fail This Moment

The absence of transparency about use of these systems weakens workers’ already limited legal protections. At-will employment means that most workers have no real basis for challenging an employment decision as unfair, whether or not ADSs are involved. After all, the collapse of unionization rates — from representing one third of the American workforce in the 1950s to 10% today — means the due-process protections that union contracts provide (just-cause provisions, grievance procedures, neutral arbitration) are unavailable to the vast majority of America’s workers.[17]

Anti-discrimination law­—one of the few ways most workers can challenge an unfair employment decision—generally requires a worker to show that discrimination occurred and to establish causation.[18] Each step becomes substantially harder when the decision-maker is opaque. Employers are not required to disclose that ADSs are in use, and the systems themselves are typically treated as proprietary — even though workers and human reviewers may need to understand how an output was generated in order to evaluate or contest it.[19]

Ongoing attacks on disparate impact doctrine sharpen this problem further.[20] This anti-discrimination framework is uniquely important for challenging algorithmic decisions. It is often difficult to show that employers using AI systems have discriminatory intent. But the AI systems often produce discriminatory outcomes by learning from biased data, sometimes relying on proxies for protected traits such as ZIP code.[21]  A disparate-impact claim is often the only viable way to challenge employment decisions made by an AI system, and the discovery it opens can be the only means by which a worker learns how an opaque model reached its decision.[22]

Requiring employers to answer for excluding qualified people without justification also creates an incentive to test for and correct bias before deployment, when less discriminatory alternatives that preserve accuracy are typically still available.[23] Yet this framework for challenging unnecessary practices that produce discriminatory results is itself under sustained assault. In 2025, the EEOC chose to halt investigations into disparate impact claims, thus limiting further the tools available to workers to address algorithmic discrimination.[24]

Workers across the economy are already living in an algorithmically managed workplace, but the legal infrastructure to protect them has not kept pace. The meaningful-human-oversight principle addresses that gap, with targeted requirements consistent with existing employer obligations and manageable by HR functions that already handle comparable processes.

2. Operationalizing the Principle of Meaningful Human Oversight

Current law in most states imposes no obligations on employers who deploy algorithmic systems to discipline, suspend, or terminate workers. The elements described below would change that, but they would not prohibit employers from using AI tools, and they do not change at-will employment. For the category of employment decisions where algorithmic management is most consequential, such as discipline and termination, this approach would require due process, including a human making the decision and the right to appeal to a human. These protections would apply to all workers subject to algorithmic management, including gig workers and independent contractors.

The principle of meaningful human oversight can be operationalized by a statute that contains the set of protections below: a human making the decision is its core, but that is only effective when paired with notice, access to the underlying data, a route to appeal, and protection from retaliation. The five elements are summarized here and detailed in the subsections that follow.

At a glance: operationalizing the principle of human oversight:

• Human ownership of the decision. A person with authority must review any automated recommendation to discipline or terminate a worker before it is final.
• Notice and transparency. Workers get advance notice that an automated system is in use, and notice after the fact when one is used for an adverse decision.
• Data access. Workers can obtain the recent data used against them and correct it.
• Appeal. Workers can appeal to a human not involved in the original decision, backed by a private right of action.
• Anti-retaliation. Workers who assert these rights are protected from retaliation.

Human Ownership of the Decision

Covered employers may not rely primarily on an ADS for discipline, termination, or deactivation decisions. When an ADS plays a role in such a decision, a human reviewer must examine the output and consider other relevant information before the decision is finalized.

This requirement operates as a check in case the ADS is missing relevant context for the decision. For example, a recent lawsuit against Meta challenged its use of AI token consumption as a factor in layoff decisions without the context that some employees were on leave during the time they were being evaluated.[25]

The requirement also addresses the well-documented problem that algorithmic systems can inherit the biases of previous decision-makers and reproduce existing patterns of discrimination.[26] When managers receive an unjust algorithmic recommendation, experimental research indicates they may follow it at a statistically significant rate rather than correcting it. [27] If a human reviewer is required to independently examine the ADS output and consider the full context, that may partially address the problem but it may not be sufficient.

Notice and Transparency

Workers must receive advance written notice before an ADS is deployed for significant employment-related decisions— including decisions about disciplining or firing. When an adverse action primarily relies on an ADS, the worker must be notified that an automated system was used and given a human to contact for more information. These requirements do not compel disclosure of proprietary algorithms.

Transparency matters because workers who receive a termination notice today typically do not know whether a human or an algorithm made the decision, what data was considered, or whether that data was accurate.

As a matter of basic fairness, pre-use notice allows workers to understand the system they are operating under, and post-use notice helps them clarify the basis for a decision and determine whether they have grounds to challenge it. Together, these requirements create an informational foundation for accountability.

Data Access Rights

Under the framework proposed here, workers would have a right to request a copy of the twelve most recent months of their own data primarily used in an ADS-driven discipline or termination decision. Employers must provide that data and inform workers, and where applicable their unions, of their right to access and correct it. Ideally, workers would also be told what data an ADS collects and be given the opportunity to consent before it is used in a covered decision. The right is limited to the worker’s own data and to the specific decisions covered by the statute.

In practice, this looks like a delivery driver whose productivity tracker recorded low output because a GPS error logged excessive idle time, or a warehouse worker whose scanner malfunctioned during an evacuation drill, requesting the underlying data and verifying its accuracy before or after challenging the decision. Under current law, neither worker has any right to see the data being used against them. Transparency about data collection is itself a form of fairness and accountability.

Appeal Process and Enforcement

Workers should also be able to appeal an ADS-driven adverse decision to a human reviewer not involved in the original determination, with access to full documentation of the ADS output used. And if employers do not comply with the requirements of meaningful human oversight, then workers need to be able to sue to hold their employer accountable. The two enforcement mechanisms are complementary: the internal appeal provides first-line resolution within the employment relationship, while a private right of action provides the backstop when that process fails or is unavailable.

States should include a private right of action backed by meaningful damages. Unfortunately, New York City’s experience demonstrates that agency enforcement alone is insufficient.[28] New York City’s Automated Employment Decision Tools law (Local Law 144 of 2021, enforced beginning July 2023), heralded as the first ADS regulation in employment, offers a cautionary example of what inadequate legislation looks like.[29] Penalties for first violations max out at $500 and subsequent violations max at $1,500 per day­—costs that are easily absorbed by large employers.[30] There is also no private right of action.[31] A 2024 analysis found that more than a year into the law’s operation, most major employers had not conducted or published the required bias audits, and enforcement remained minimal.[32]  The lack of enforcement also reflects a broader national pattern in which the agencies responsible for enforcing worker protection statutes are chronically underfunded and increasingly outmatched by the scale and speed at which algorithmic systems operate.[33]

But the efficacy of that law was also seriously undercut by its narrow scope: excluding even termination decisions.[34] The law required only that employers notify candidates at least ten business days before using such a tool to screen candidates for hire or employees for promotion, and only those that substantially assist or replace discretionary decision-making. That narrow scope is a large part of why it does little to close the human-oversight gap this brief addresses.

The federal No Robot Bosses Act (H.R. 6371) provides a better model.[35] It includes statutory damages of $5,000–$20,000 per violation (up to $40,000 for willful violations) plus attorneys’ fees, prohibition of predispute arbitration agreements and class action waivers, and State Attorneys General authority to bring parens patriae actions.[36] Statutory penalties are also appropriately indexed to inflation so that their deterrence value does not erode over time.[37]

Anti-Retaliation Protection

Workers who assert rights under the law are protected from discharge, discipline, or other retaliation, enforceable through the relevant state labor agency (for example, a state Labor Commissioner or Attorney General) and by workers themselves. Without enforceable anti-retaliation protections backed by a private right of action, the rights above become practically unenforceable for workers in vulnerable employment situations. The power asymmetry between employer and worker makes voluntary compliance insufficient on its own.

Background Checks: A Similar Requirement. Similar due process requirements are already part of federal law under the Fair Credit Reporting Act (FCRA). The FCRA requires employers who use third-party background reports to provide stand-alone written disclosures before obtaining a report, secure written authorization from the applicant, and deliver pre-adverse action notice — including a copy of the report—before rejecting a candidate based on its contents.[38] The FCRA’s notice and disclosure obligations apply to all employers regardless of size.[39] So even the smallest employers should be familiar with the basic practice of notifying workers and applicants when information is used against them.

The FCRA also illustrates what a meaningful review step looks like in practice. Before an employer may take final adverse action based on a third-party background report, it must give the worker a copy of the report and a written summary of rights, and then wait a reasonable period before finalizing the decision, so the worker can respond and correct inaccuracies.[40] The waiting period is critical. It rests on the premise that a report assembled at a distance from the worker may be wrong, and that the worker is the person best positioned to say so. A meaningful-human-oversight requirement for ADS-driven discipline or termination would work the same way: notice, a genuine opportunity to correct the record, and a human decision-maker who must consider the response before the action becomes final.

The FCRA’s implementation record also shows which design choices matter. Employers rarely run the process themselves. They buy it from the screening company, which scores the report against a pre-loaded “adjudication matrix” of disqualifying offenses and issues the required notices on an automated clock—one major vendor defaults to seven calendar days, though advocates report that correcting an inaccurate report takes at least two weeks.[41] Compliance breaks down at the notice step: employers and screening companies paid more than $325 million to settle FCRA claims in the decade ending in 2019, and Amazon itself settled such a claim for $5 million in 2018, brought by an applicant denied a job on the basis of a report he was never shown.[42]

This record identifies what a human-oversight statute should do differently than the FCRA: set a defined minimum response period rather than a “reasonable” one, require that notice reach the worker before an adverse determination goes to the decision-maker, and include a private right of action, since federal enforcement here has focused almost entirely on the reporting companies rather than the employers who act on their reports.[43]

3. Existing Models: What Implementation Looks Like

Critics may argue that these requirements will force companies to redesign their operations from the ground up. But we actually have recent examples of this kind of regulation in action—not from ADS regulation per se, but from legislation targeted at deactivation of app-based gig workers. We also have examples of non-union employers voluntarily providing their employees with this kind of due process. This record indicates that the required processes are manageable for employers.

Seattle and Minnesota: Gig-Worker Deactivation Laws

Gig workers, including rideshare drivers, delivery couriers, and domestic workers, face the most acute version of algorithmic management. As some of the most precarious workers, they are the canaries in the coal mine. When an app-based worker is deactivated, they lose access to the platform entirely and often their primary income source instantly, with no notice and no avenue for appeal. The decision is typically made by an automated system processing complaints, ratings, or performance data. In many cases, workers are not told why they were deactivated or even which data triggered the decision.

Seattle became the first jurisdiction in the country to regulate app-based deactivations with Ordinance 126878 in 2023.[44] The ordinance bars deactivations based solely on customer ratings or unsubstantiated complaints and requires companies to demonstrate a policy violation by a preponderance of the evidence, applied consistently and proportionately.[45] Companies must provide 14 days’ advance written notice with access to all records relied upon.[46] Workers may contest the deactivation within 90 days, and the company must respond within 14 days through an individual with authority to reinstate the worker.[47] The Seattle City Council’s legislative findings documented that consumer-sourced rating systems are highly likely to be influenced by bias on the basis of race or ethnicity, and that many platforms do not have processes to substantively reconsider a deactivation based on a case-by-case human review.[48]

Platform companies quickly challenged the ordinance.[49] Uber and Instacart argued it violated their First Amendment rights by compelling disclosure of deactivation policies.[50] In March 2026, the Ninth Circuit affirmed the denial of a preliminary injunction against the ordinance in Uber Technologies, Inc. v. City of Seattle, holding that it regulates nonexpressive conduct, specifically the unwarranted deactivation of worker accounts, and that any incidental burden on the companies’ speech is not constitutionally significant.[51] The court noted that the ordinance requires disclosure only to directly affected workers, not to the general public.[52] Seattle’s platform operations have continued without disruption.

Minnesota followed in 2024 with comparable protections for ride-hail drivers under Minn. Stat. § 181C.04. The Minnesota law requires plain-language written policies, written notice of deactivation stating the reason and the driver’s right to appeal, at least 30 days to appeal, during which the driver may submit additional information that the company must consider before responding within 15 days, and independent driver advocacy organizations contracted to provide workers with free representation.[53] Where a purely technical error causes an unintentional deactivation, the driver is entitled to compensation for the period they were locked out, up to 21 days.[54]  Minnesota’s experience points the same way: the deactivation protections took effect in December 2024 as part of a statewide compromise that resolved an earlier standoff over driver pay, and both Uber and Lyft continued operating across the state—now contracting with independent driver-advocacy organizations to help drivers appeal deactivations, as the law requires.[55]

New York City

New York City sought to extend this model in 2026. After overriding Mayor Adams’s veto, the City Council enacted Local Law 52, requiring just cause, fourteen days’ advance notice, progressive discipline, and a challenge process before a high-volume for-hire vehicle service may deactivate a driver.[56] But Uber and Lyft challenged the law as unconstitutional, and one week before the law took effect, the Southern District of New York preliminarily enjoined it in full.[57]  As of this writing, the decision is on appeal by the City.

The ruling is narrower than it first appears, and its reasoning is instructive for drafters. It rested entirely on the Constitution’s Contracts Clause, which restricts legislation that impairs existing contractual obligations. That constraint has little purchase on the human oversight principle as applied to the general workforce. It may have force where a statute arguably changes the terms of an existing platform-worker agreement, as with Uber and Lyft.

Two unique features of the law also drove the analysis. First, the law reached backward, permitting drivers to petition for reinstatement based on deactivations occurring up to seven years before enactment.[58] Second, the court found that the legislative record did not address the law’s consequences for the broader public—passenger and pedestrian safety, service costs, or public expenditures—and concluded on that record that the law was likely enacted for the benefit of a narrow class rather than in service of a general social or economic problem.[59] These are the kinds of issues that could be addressed through careful drafting and a more developed legislative record for future laws.

Deactivation-protection laws for app-based workers.

Protection	Seattle (2023)	Minnesota (2024)	New York City (2026)
Grounds for deactivation	May not deactivate based solely on customer ratings; requires just cause established by a preponderance of the evidence.	Platform must maintain a plain-language written policy stating the grounds on which a driver may be deactivated.	Just cause required, with progressive discipline before deactivation.
Notice	14 days’ written notice before deactivation.	Written notice at the time of deactivation, stating the reason and the driver’s right to appeal; a warning is required before deactivation unless the conduct is defined as serious misconduct.	14 days’ advance written notice.
Access to data	Access to the records the platform relied upon in making the decision.	Access to the records supporting the stated reason for deactivation.	Access to the data underlying the decision.
Appeal and human review	90-day window to contest the deactivation, with review by a human.	At least 30 days to appeal; the company must respond within 15 days. Independent driver advocacy organizations provide free representation, and external reviewer permitted.	Challenge process available before deactivation takes effect.
Status	In effect.	In effect.	Preliminarily enjoined in July 2026; decision is on appeal.

The Broader Workforce: Existing Employer Practices Provide a Model

What Seattle and Minnesota require of platform companies is what the human oversight principle would require of all employers for their most consequential decisions, and it is comparable to what some companies already do voluntarily.

IBM’s Open Door process, for example, allows any employee to request that a senior manager investigate a workplace decision, or to convene a randomly selected panel of three employees and two managers to review a challenged determination.[60] IBM has maintained this process across decades for a variety of employment decisions.[61] The approach is simple: it requires only reviewers with authority, access to relevant data, and a process through which the worker can be heard.

FedEx offers a second long-standing example, and one that also reaches a largely non-union workforce. Its Guaranteed Fair Treatment Procedure lets a worker contest an adverse decision, including a termination, through successive levels of review. Like IBM, FedEx adopted the procedure voluntarily. Both are simply how some large employers have chosen, for decades, to put a human check on consequential personnel decisions even absent any collective-bargaining or legal obligation to do so. [62]

HR departments at companies also routinely handle some comparable review processes. Besides background checks under the FCRA, accommodation requests under the ADA and Pregnant Workers Fairness Act (“PWFA”), medical leave certifications under the FMLA, and internal grievances all require individual case review, documentation, and a response within defined timeframes. These examples illustrate that human review of employment decisions is operationally familiar. When these laws were debated, employer groups raised the same objections now leveled at requirements for algorithmic employment decisions.

4.  Addressing Objections

Opposition to the principle of human oversight in algorithmic employment decisions, led by business associations in jurisdictions where it has been debated, raises several objections. Each points toward calibration in how the principle is implemented, not toward abandoning it.

Objection 1: These requirements are too broad, impose excessive compliance costs, and will chill innovation in AI tools that could actually reduce bias.

A coalition led by the California Chamber of Commerce argued that laws like California’s 2025 proposal (SB 7) “broadly target businesses of all sizes, across every industry, and regulate even low-risk applications of automated decision systems (ADS) or where there is human involvement in a decision in addition to the ADS,” warning that requirements would be costly and produce a “chilling effect on the technology and make it that much harder to develop the very tools that can help combat bias in decision making.”[63]

These arguments have significant problems.

Response:

• “Chills innovation.” The argument is that regulation will inhibit the development of AI tools that reduce bias, but the current unregulated environment has produced tools that demonstrably reproduce bias, as documented in the academic literature on algorithmic prediction.[64] Accountability requirements are what create market pressure for better, less biased tools. As to the general “chill innovation” claim, it is difficult to believe that such a modest increase (if any) in the cost of using these tools would significantly impact adoption and therefore innovation.
• Cost. Terminations and serious disciplinary actions are not everyday events. Calibrating the requirement to high-stakes decisions (discipline, termination) limits both the scope of the obligation and the frequency of its invocation.

Furthermore, HR infrastructure to enable regulatory compliance already exists. When Congress debated the FMLA, employer groups warned that compliance would be costly and disruptive; studies after enactment found little or no effect on productivity, profitability, or costs, and benefits, such as reduced turnover, that offset the expense.[65] Similarly, when Congress passed the ADA in 1990, some warned it would impose undue administrative burdens. But many states had already enacted their own disability antidiscrimination laws, and a substantial number already required reasonable accommodation—so for a large share of employers the ADA extended review and documentation practices they were already running rather than creating them.[66] Both statutes require individualized review, documentation, and response within defined timeframes.

More recently, when Congress passed the Pregnant Workers Fairness Act and the PUMP for Nursing Mothers Act together in late 2022, employers again raised burden-and-cost objections, yet most complied by extending the accommodation and documentation processes their HR departments already ran under the ADA and FMLA.[67] In each case employers were able to adapt.

• Breadth. The breadth concern is addressable through scope. The California Chamber of Commerce itself conceded that it appreciated “concerns over employees being disciplined or terminated solely based on automated tools”, and its objection was that SB 7 was not tailored to those scenarios.[68] That is an argument for precise scoping, not for inaction. A framework focused on high-stakes decisions like discipline and termination—the decisions that most directly threaten a worker’s livelihood—addresses the concern while preserving the core accountability requirement. Indeed, the 2026 version of the bill in California is narrower and uses such a framework.[69]

Objection 2: Certain automated decision systems should be permitted to drive decisions, particularly where safety is involved or where customer feedback is a legitimate performance metric.

Employer groups have argued that some ADS applications serve legitimate purposes that blanket human-oversight requirements would compromise. They assert that ADS can detect workplace safety violations and that requiring independent human corroboration before acting is overly cumbersome.[70] They have also argued that customer ratings are a meaningful performance indicator, especially in roles where supervisors “are not always present,” and that the principle of not relying primarily on customer ratings is overbroad.[71]

These arguments identify an implementation question, but not a reason to abandon the principle of meaningful human oversight.

Response:

• Safety. Safety-triggered actions and performance-driven terminations are categorically different, and legislation can reflect that distinction. A rideshare platform that receives a serious safety complaint has every reason to suspend a driver’s access immediately, and nothing in a human oversight requirement should prevent that. What the requirement addresses is what would happen next, like whether a human being reviews the underlying data and considers the full context before that suspension becomes a permanent deactivation. Speed in the face of a safety concern and accountability in the decision to permanently end someone’s livelihood are not in conflict. A framework focused on discipline and termination can protect both values.
• Customer Ratings. Customer ratings can be a meaningful performance indicator. But as the Seattle City Council and researchers have found, consumer-sourced rating systems are highly likely to be influenced by bias on the basis of race or ethnicity.[72] Seattle’s solution was not to ban the use of customer feedback but to prohibit deactivation based solely on ratings. Platforms can still use ratings as one input among others.
• Due Process. Fundamentally, the meaningful human oversight principle is about due process. Algorithmic outputs are inferences drawn from data that may be inaccurate, biased, or collected under conditions the worker could not control. Human review is not about slowing down legitimate enforcement. It is about ensuring that the data being acted upon is actually what it appears to be, and that the decision reflects the full legal and factual context a worker is entitled to have considered.

Objection 3: An appeal requirement imposes impracticable burdens on small employers and undermines at-will employment.

Opponents have argued that an appeal requirement is “not necessary to ensuring a human is in the loop,” that it is “contrary to at-will employment,” and that it would “grind workplaces to a halt and create unnecessary hurdles to everyday decisions.”[73] They also asserted that small employers, particularly those in which a single manager handles all personnel decisions, could not provide a reviewer “not involved at all in the original decision.”[74]

These arguments are easily resolved.

Response:

• At-Will. The at-will argument misunderstands what is being required. At-will employment means an employer can fire a worker for any lawful reason, even if not job-related. The human oversight principle does not change that. An employer can still reach the same decision after a human review. The employer does not have to give a good reason; it must simply give a reason and give the worker a chance to respond. This is a process requirement, not a substantive constraint on the termination decision itself.
• Unnecessary Hurdles. The “grind workplaces to a halt” argument is vastly overstated. Terminations and serious disciplinary actions are not everyday occurrences for most employers, which is in part why the burden is modest. And the window for workers to appeal should be relatively short, as it is in most such laws. That process does not grind workplaces to a halt, and giving workers due process—particularly when AI is involved in making a consequential employment decision—is far from an “unnecessary hurdle.”
• Small Employers. For smaller employers with limited HR capacity, the solution Minnesota found workable for its deactivation law applies equally here: designation of an external reviewer, such as a contracted HR professional or an industry association service.[75] Outsourced human-resources providers — professional employer organizations and payroll-and-HR firms such as Paychex, ADP, or TriNet—already supply exactly this kind of third-party review and administration to small employers, and could readily absorb the modest additional task of an independent ADS review. The ADA and FMLA also demonstrate that tiered obligations based on employer size are another possibility for balancing small-business concerns with worker protections, though there are of course serious tradeoffs in worker protection with this approach.
• Appeal Not Necessary. The argument that appeal rights are “not necessary to ensure a human is in the loop” might be reasonable except for the research indicating that when managers receive an unjust algorithmic recommendation, they often follow it rather than correcting it.[76]  Requiring human ownership of the initial decision should help address this issue, but it will be difficult to detect when managers simply rubber-stamp the AI recommendation. When a human reviewer on appeal must document their decision and provide a reason in writing to the worker, there should be fewer arbitrary decisions. And knowing that a human may review an appeal should incentivize the initial manager to not simply defer to the AI.

Conclusion

The warehouse worker who received an automated termination notice had no way to know whether a human reviewed her case, no right to the data used against her, and no structured process to contest the decision. Under a human oversight requirement, she would have had all three.

The meaningful-human-oversight principle does not prevent employers from using AI. Instead, it prevents using AI to eliminate accountability from the most consequential decisions employers make about workers’ lives. It asks for a human reviewer, access to data, and due process. Existing law like FCRA for background checks already has similar requirements. Companies like IBM and FedEx have provided due process to workers for decades, even outside a union context. And Seattle and Minnesota have enacted comparable requirements for app-based gig workers, with platform operations able to adapt.

The principle of meaningful human oversight commands broad public support at a time of high levels of worker fear around how AI is going to affect their lives. The workers who need this protection are already living in an algorithmically managed workplace. They are not waiting for the technology to arrive, but they are waiting for the law to catch up.

Acknowledgements

This brief benefited from the thoughtful review and comments of Mary Beech, Tanya Goldman, Lauren McFerran, Matt Scherer, Ridhi Shetty, Sara Steffens, and Crystal Weise. Their insights materially improved the analysis. Thanks also to Abby Frerick for all her terrific work on this. Any errors, and the views expressed, are NIWR’s alone.

Appendix

Legislation

Three kinds of legislation now address algorithmic management, arising from different starting points but converging on the same core protections. Human-oversight bills place a person with authority between an automated system and a consequential employment decision. Gig-worker deactivation laws focus on the platform economy, and a newer set of ADS and anti-discrimination laws regulates automated decision systems primarily through transparency. The table below provides examples from each category and scores each law against the five elements of the framework set out in Section II. This reflects the state of the law as of July 2026.

Columns below track the five framework elements: human ownership of the decision, notice and transparency, data access, appeal (backed by a private right of action, “PRA”), and human ownership of the decision.

✓ present   ◑ partial or qualified   — absent. Columns track the five framework elements: human oversight, notice and transparency, data access, appeal (backed by a private right of action, “PRA”), and anti-retaliation.

Law	Jurisdiction / status (As of July 2026)	Human ownership	Notice	Data access	Appeal / PRA	Anti-retaliation
Human-oversight bills (No Robo Bosses model)
SB 9471	California — passed Senate May 2026; pending in Assembly	✓	◑	✓	—	✓
H.R. 63712	Federal/House — introduced December 2025	◑	✓	✓	✓	✓
S. 48333	Federal/Senate — introduced June 2026	✓	◑	◑	◑	✓
Gig-worker deactivation laws
Local Law 524	New York City — enacted 2026, enjoined	◑	✓	✓	✓	✓
Ordinance 1268785	Seattle — enacted 2023, in effect	◑	✓	✓	✓	✓
§ 181C.046	Minnesota — enacted 2024, in effect	—	✓	◑	◑	—
ADS / anti-discrimination laws
SB 24-2057	Colorado — amended 2026 (eff. Jan. 2027)	—	✓	✓	◑	—
HB 37738	Illinois — enacted 2024, in effect	—	◑	—	◑	◑
SB 5 (CART Act)9	Connecticut — enacted 2026 (eff. Oct 1, 2026)	—	◑	—	—	—

1. Human review before discipline, termination, or deactivation; advance notice; data access; appeal process; private right of action.

2. Notice; data access; formal appeal; private right of action with $5,000–$20,000 statutory damages per violation; no forced arbitration. Required pre-deployment testing and post-use impact assessment, as well as ability of worker to opt out of being managed by an ADS

3. Key differences from House bill include stronger human ownership requirement (decision may not be “predominantly” made by ADS as opposed to “exclusively” in House bill), but no post-use notice, required appeal, or data access.

4. Just cause, 14 days’ notice, progressive discipline, data access, and a challenge process before deactivating appbased drivers.

5. Bars deactivation on ratings alone; just cause by a preponderance of the evidence; 14 days’ notice and record access; 90-day human-review window.

6. Written deactivation policies and reasons; 30-day appeal window with a 15-day company response deadline; platform-contracted driver advocacy organizations provide free representation; external reviewer permitted where none is available in-house.

7. Notice, appeal-to-human-review after adverse decisions, and data-correction rights, but no predecision human-oversight mandate. AG-only enforcement, no private right of action.

8. Amends the Illinois Human Rights Act: notice of AI use required and AI-driven discrimination barred (including ZIP-code proxies), but no oversight, data-access, or appeal mechanism. Enforced through the IHRA, which permits a civil action in circuit court.

9. Pre-use notice; use of an automated tool is no defense to a discrimination claim. No pre-decision oversight mandate and no private right of action for the notice provisions (AG-only).

---

Endnotes

[1] Kate Gibson, Amazon Under Fire for Software That Recommends Firing Workers, CBS News (Apr. 26, 2019), https://www.cbsnews.com/news/amazon-under-fire-for-software-that-recommends-firing-workers/.

[2] Spencer Soper, Fired by Bot at Amazon: It’s You Against the Machine, Bloomberg, June 28, 2021, https://www.bloomberg.com/news/features/2021-06-28/fired-by-bot-amazon-turns-to-machine-managers-and-workers-are-losing-out.

[3] Alexander Hertel-Fernandez, Estimating the Prevalence of Automated Management and Surveillance Technologies at Work and Their Impact on Workers’ Well-Being: Evidence from a New National Survey and Implications for U.S. Federal Policy 10 (Wash. Ctr. for Equitable Growth, Oct. 2024), https://equitablegrowth.org/wp-content/uploads/2024/10/workplace-surveillance-report50.pdf [hereinafter Hertel-Fernandez].

[4] 42 U.S.C. § 2000e-2.

[5] AI and Work Survey: Conducted April 14–22, 2026, David Binder Research for the AFL-CIO (May 2026), https://aflcio.org/sites/default/files/2026-05/DBR_AFLCIO_AI_Research_Memo.pdf (reporting 95% support among the 1,164 surveyed voters who are also workers).

[6] In this brief, we will sometimes use “AI” to refer to the broader category of software of which Automated Decision Systems (ADSs) are a subset.  

[7] Andrea Meyer, Manager Insights: Your Secret AI Coaching Superpower, Centrical (archived Mar. 16, 2026), https://perma.cc/K5YH-YF2S.

[8] WorkTime Homepage, WorkTime (archived Mar. 16, 2026), https://perma.cc/YMY4-MMZY.

[9] Predictive Assessments Backed by Science, Harver (archived Mar. 16, 2026), https://perma.cc/U7A2-UA5U.

[10] Perceptyx Homepage, Perceptyx (archived Mar. 16, 2026), https://perma.cc/9NFN-W6VQ; Bradley Wilson, Influence the Future with Predictive Analytics in HR, Perceptyx Blog (May 8, 2019)(archived Mar. 16, 2026), https://perma.cc/JT9M-GMWA.

[11] Kathlyn Chua, Revolutionizing Pre-Employment Screening: The Power of AI in Document Fraud Detection, Resistant AI Blog (May 22, 2024)(archived Mar. 16, 2026), https://perma.cc/SE9A-63AH.

[12] Arvind Narayanan & Sayash Kapoor, AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference 33 (2024).

[13] Juliana Kim, Amazon Manipulated Injury Data to Make Warehouses Appear Safer, a Senate Probe Finds, NPR (Dec. 16, 2024), https://www.npr.org/2024/12/16/nx-s1-5230240/amazon-injury-warehouse-senate-investigation; Lauren Kaori Gurley, Internal Documents Show Amazon’s Dystopian System for Tracking Workers Every Minute of Their Shifts, Vice (June 2, 2022), https://www.vice.com/en/article/internal-documents-show-amazons-dystopian-system-for-tracking-workers-every-minute-of-their-shifts/; Gibson, supra note 1.

[14] See The New HotSOS Housekeeping Experience, Amadeus Hospitality (archived Mar. 16, 2026) https://perma.cc/9SWS-34MF; Why Most Security Providers Are Still Under-Adopting Automation — and How That’s a Competitive Gap, Trackforce (Jan. 14, 2026) (archived Mar. 16, 2026) https://perma.cc/A67J-GY9H. For a detailed study of algorithmic management in hospitality, developed with UNITE HERE, see “Unseen”: Algorithmic Management and Hospitality Workers, Proceedings of the 2023 ACM Conference on Computer-Supported Cooperative Work (2023), https://dl.acm.org/doi/pdf/10.1145/3563657.3596018.

[15] Hertel-Fernandez, supra note 3, at 12.

[16] Id. at 13.

[17] Laura Feiveson, Labor Unions and the U.S. Economy, U.S. Dep’t of the Treasury (Aug. 28, 2023), https://home.treasury.gov/news/featured-stories/labor-unions-and-the-us-economy.

[18] 42 U.S.C. § 2000e-2.

[19] Workers have begun to challenge the opacity of algorithmic personnel decisions directly. See, e.g., Barbara Ortutay and Alexandra Olson, 26 Meta Workers Sue, Alleging AI-Driven Layoff Picks Hits Workers on Medical and Parental Leave, Associated Press (July 15, 2026), https://apnews.com/article/meta-lawsuit-workers-target-ai-layoffs-leave-019fb9c7fdc09167e91547546bce5be8.

[20] See Disparate Impact Civil Rights Claims: A Crucial Tool Under Attack, Democracy Forward, National Institute for Workers’ Rights, The Leadership Conference on Civil and Human Rights, and Legal Defense Fund (2025), https://niwr.org/2025/11/03/disparate-impact/.

[21] See Chiraag Bains, When Machines Discriminate: The Critical Role of Disparate Impact in AI Accountability 21–24, The Leadership Conference on Civil and Human Rights, https://civilrights.org/disparate-impact-ai/.

[22] See Jason Solomon and Abby Frerick, “Congress is About to Hand Corporate America a License to Discriminate,” Tech Policy Press, June 30, 2025.

[23] Bains, supra note 21, at 25–26.

[24] Exec. Order No. 14281, Restoring Equality of Opportunity and Meritocracy, §§ 4, 6(a), 90 Fed. Reg. 17537, 17538 (Apr. 28, 2025) (directing agencies to “deprioritize enforcement of all statutes and regulations to the extent they include disparate-impact liability” and directing the EEOC Chair to assess all pending matters resting on that theory); Rebecca Klar, EEOC to Close Workers’ Disparate Impact Job Bias Charges, Bloomberg Law Daily Lab. Rep. (Sept, 19, 2025) (reporting an internal EEOC memorandum directing closure of nearly all charges alleging only disparate impact).

[25] See Complaint, Does 1-26 v. Meta Platforms, Inc., Case 3:26-cv-07122-WHO (N.D. CA, filed 7/13/2026).    

[26] Maryam Ghasemaghaei & Nima Kordzadeh, Understanding How Algorithmic Injustice Leads to Making Discriminatory Decisions: An Obedience to Authority Perspective, 61 Info. & Mgmt. 103921, 3 (2024), https://doi.org/10.1016/j.im.2024.103921.

[27] Marcin Bartosiak & Artur Modliński, Fired by an Algorithm? Exploration of Conformism with Biased Intelligent Decision Support Systems in the Context of Workplace Discipline, 27 Career Dev. Int’l 601, 602 (2022), https://doi.org/10.1108/CDI-06-2022-0170.

[28] Kenrick Sifontes, Enforcement of Local Law 144 – Automated Employment Decision Tools, N.Y. Office of the State Comptroller (Dec. 2, 2025), https://www.osc.ny.gov/state-agencies/audits/2025/12/02/enforcement-local-law-144-automated-employment-decision-tools.

[29] N.Y.C. Admin. Code §§ 20-870 to 20-874 (Local Law 144 of 2021).

[30] N.Y.C. Admin. Code § 20-872(a)–(b).

[31] N.Y.C. Admin. Code § 20-874 (preserving existing human rights claims but creating no private right of action under the AEDT law).

[32] Sifontes, supra note 28.

[33] Worker Protection Agencies Need More Funding to Enforce Labor Laws and Protect Workers, Econ. Pol’y Inst. (July 29, 2021), https://www.epi.org/blog/worker-protection-agencies-need-more-funding-to-enforce-labor-laws-and-protect-workers/.

[34] N.Y.C. Admin. Code §§ 20-870 (defining “employment decision” to mean only “screen candidates for employment or employees for promotion within the city”).

[35] No Robot Bosses Act, H.R. 6371, 119th Cong. (2025), https://www.congress.gov/bill/119th-congress/house-bill/6371/text.

[36] Id. § 7(a)(3)(A)–(B), (e)(1), (f).

[37] Id. § 7(a)(3)(B)(vi).

[38] 15 U.S.C. § 1681a(b) (defining “person” to include any individual, partnership, corporation, trust, estate, cooperative, association, government, or governmental subdivision); 15 U.S.C. § 1681b(b)(2)–(3) (imposing disclosure and adverse action obligations on any “person” who uses a consumer report for employment purposes).

[39] 15 U.S.C. § 1681a(b) (defining “person” to include employers of any size); the FCRA contains no small-employer exemption.

[40] 15 U.S.C. § 1681b(b)(3)(A) (requiring the employer to provide a copy of the report and a written summary of rights before taking adverse action). The reinvestigation duty for disputed information runs against the consumer reporting agency, not the employer. 15 U.S.C. § 1681i(a).

[41] See, e.g., HireRight, Adjudication Service, https://www.hireright.com/services/adjudication (automatic adjudication of “clear” reports); Bchex, Insight+ (predefined rules “applied to every background check, eliminating variability between reviewers”); Nat’l Consumer Law Ctr., Broken Records Redux: How Errors by Criminal Background Check Companies Continue to Harm Consumers Seeking Jobs and Housing 12–13, 31–32 (Dec. 2019), https://www.nclc.org/wp-content/uploads/2022/09/report-broken-records-redux.pdf (describing a screening company’s seven-calendar-day default waiting period and automatic transmission of the post-adverse-action notice, and reporting that correcting an inaccurate report takes at least two weeks on average) [hereinafter Broken Records Redux]. The FCRA specifies no minimum waiting period, and the FTC has said only that it must be “reasonable.” 15 U.S.C. § 1681b(b)(3); Fed. Trade Comm’n, 40 Years of Experience with the Fair Credit Reporting Act: An FTC Staff Report with Summary of Interpretations (July 2011).

[42] Megan Cerullo, What Everyone Should Know About Employer Background Checks, CBS MoneyWatch (June 28, 2019), https://www.cbsnews.com/news/what-job-candidates-should-know-about-employer-background-checks/ (reporting Good Jobs First data: roughly $174 million paid by employers and $152 million by background check companies over the preceding decade, and describing Amazon’s 2018 settlement of claims that the applicant “was deprived of any opportunity to review the information in the report and discuss it with Defendant before he was denied employment”); see also Broken Records Redux, supra note 41, at 31 (“the first breakdown of consumer protection laws often occurs because employers fail to comply with the FCRA’s notice requirements”).

[43] Federal enforcement has been directed at the reporting companies rather than the employers who use their reports. See Consumer Fin. Prot. Bureau, CFPB Takes Action Against Two of the Largest Employment Background Screening Report Providers for Serious Inaccuracies (Oct. 29, 2015) ($10.5 million in consumer relief and a $2.5 million penalty; nearly 70 percent of criminal history disputes filed with one provider between 2010 and 2014 produced a change or correction); Consumer Fin. Prot. Bureau, Sterling Infosystems, Inc. (Nov. 2019) ($6 million in relief, $2.5 million penalty). Federal guidance has also receded: CFPB Circular 2024-06, which concluded that third-party algorithmic worker scores are consumer reports subject to the FCRA, was withdrawn on May 12, 2025. Interpretive Rules, Policy Statements, and Advisory Opinions; Withdrawal, 90 Fed. Reg. 20,084 (May 12, 2025).

[44] Seattle Ordinance 126878 (2023), codified at Seattle Mun. Code ch. 8.40 (App-Based Worker Deactivation Rights Ordinance; effective Jan. 1, 2025).

[45] Seattle Mun. Code § 8.40.050(A)(2), (A)(4), (A)(5).

[46] Seattle Mun. Code § 8.40.070(A).

[47] Seattle Mun. Code § 8.40.060(B).

[48] Seattle City Council, Council Bill 120580, An Ordinance Relating to App-Based Worker Labor Standards (2023), https://seattle.legistar.com/ViewReport.ashx?M=R&N=Text&GID=393&ID=5460628; App-Based Workers Speak: Studies Reveal Anxiety, Frustration, and a Desire for Good Jobs, National Employment Law Project, 12, 18 (Oct. 2021), https://www.nelp.org/app/uploads/2021/11/App-Based-Workers-Speak-Oct-2021-1.pdf.

[49] Uber Techs., Inc. v. City of Seattle, No. 25-231, 2026 WL 603711 (9th Cir. Mar. 4, 2026).

[50] Id. at *3.

[51] Id. at *4-*6.

[52] Id. at *7.

[53] Minn. Stat. § 181C.04, subd. 5(b)–(c), 4(a)–(b) (2024).

[54] Minn. Stat. § 181C.04, subd. 5(e) (2024).

[55] Minn. Stat. §§ 181C.01–.10; id. § 181C.04, subd. 4 (“A TNC must contract with a driver’s advocacy organization to provide services to drivers under this section.”); Minn. Dep’t of Labor & Indus., Protections for Transportation Network Company (TNC) Drivers, https://www.dli.mn.gov/tnc (“Dec. 1, 2024: Minimum pay rate, pay rate notification to drivers, deactivation process and driver advocacy organization requirements began.”). On the legislative compromise, see Brian Basham, House Lawmakers Pass Agreement That Would Keep Uber, Lyft Operating in Minnesota, Minn. House of Representatives Session Daily (May 19, 2024), https://www.house.mn.gov/sessiondaily/Story/18408 (statewide bill followed the companies’ threat “to cease operations after the Minneapolis City Council voted to require a raise” for drivers; “[o]fficials representing Uber and Lyft assured the conference committee Sunday that the two companies would continue to operate in Minnesota”).

[56] N.Y.C. Local Law 52 of 2026; N.Y.C. Admin. Code §§ 20-1281 to 20-1290. The Council passed Int. No. 276-A on December 18, 2025; Mayor Adams vetoed it on December 31, 2025; and the Council overrode the veto on January 29, 2026. The law was to take effect July 28, 2026.

[57] Uber Techs., Inc. v. City of New York, No. 1:26-cv-04893-GHW (S.D.N.Y. July 21, 2026) (order granting preliminary injunction). The court declined to sever, finding no textual division separating the law’s application to pre-enactment contractual rights from its application to later-formed rights, and enjoined Local Law 52 in full. Slip op. at 50–52.

[58] N.Y.C. Admin. Code §§ 20-1281, 20-1283(a)–(b).

[59] Uber Techs., Inc. v. City of New York, supra note 57, slip op. at 32–38. This aspect of the decision warrants close attention. The “legitimate public purpose” inquiry asks whether a law pursues “a broad societal goal” rather than “the interests of a narrow class.” Conn. State Police Union v. Rovella, 36 F.4th 54, 63 (2d Cir. 2022). The court measured the protected class against the general population, estimating roughly 872 annual deactivations, or about 0.01% of New York City residents, and treated that ratio as evidence that the law served a favored group rather than the public. Slip op. at 34–35. Applied broadly, that reasoning would put a great many worker protection statutes at risk. The Second Circuit has not adopted that framing: it has upheld laws benefiting defined groups where the record tied the regulated conduct to a broader social or economic problem. See Melendez v. City of New York, 16 F.4th 992, 1036–38 (2d Cir. 2021); Buffalo Teachers Fed’n v. Tobe, 464 F.3d 362, 368–69 (2d Cir. 2006). The decision is a preliminary ruling by a single district court and is subject to appeal.

[60] Kalola v. Int’l Bus. Machines Corp., No. 13-CV-7339 (VB)(LMS), 2017 WL 3394115, at *2 (S.D.N.Y. Feb. 28, 2017), report and recommendation adopted, 2017 WL 3381896 (S.D.N.Y. Aug. 4, 2017) (“IBM has an ‘Open Door’ process whereby any employee can request that a senior manager conduct an investigation into the employee’s complaint. An employee can also request a ‘Panel Review’ process whereby a panel consisting of three employees and two managers (selected randomly), question the employee and his or her manager about a management decision which affects the employee.”).

[61] O’Brien v. Int’l Bus. Machines, Inc., No. 06-4864 (FLW), 2009 WL 806541, at *8 (D.N.J. Mar. 27, 2009) (use of the policy following a poor performance rating); Raymond v. Int’l Bus. Machines Corp., 954 F. Supp. 744, 753 (D. Vt. 1997) (use of the policy following a discharge).

[62] See Warner v. Federal Express Corp., 174 F. Supp. 2d 215, 218 (D.N.J. 2001) (describing the Guaranteed Fair Treatment Procedure, set forth in the FedEx Employee Handbook, as a three-step process of management review, officer review, and executive review, available to challenge adverse employment decisions including termination).

[63] Employment: Automated Decision Systems: Hearing on S.B. 7 Before the Assemb. Comm. on Privacy & Consumer Prot., 2025–2026 Leg., Reg. Sess., 13 (Cal. 2025) [hereinafter S.B. 7 Hearing].

[64] See Narayanan & Kapoor, supra note 12, at 33.

[65] Donna R. Lenhoff & Lissa Bell, Government Support for Working Families and for Communities: Family and Medical Leave as a Case Study 3, Nat’l P’ship for Women & Families, https://nationalpartnership.org/wp-content/uploads/2023/02/fmla-case-study-lenhoff-bell.pdf.

[66] Inst. of Med., The Future of Disability in America 273–74 (Marilyn J. Field & Alan M. Jette eds., 2007) (“many states had laws against discrimination on the basis of disability before the ADA came into effect, so we cannot assume that the implementation of Title I represented a change in the rules for all employers and workers”); Christine Jolls & J.J. Prescott, Disaggregating Employment Protection: The Case of Disability Discrimination (Nat’l Bureau of Econ. Rsch., Working Paper No. 10740, 2004) (classifying states as having no comparable protection, “protection without accommodation,” or “ADA-like” law that both barred discrimination and required reasonable accommodation).

[67] Pregnant Workers Fairness Act, Pub. L. No. 117-328, div. II, 136 Stat. 4459 (2022) (codified at 42 U.S.C. §§ 2000gg to 2000gg-6); Providing Urgent Maternal Protections for Nursing Mothers Act (PUMP Act), Pub. L. No. 117-328, div. KK, 136 Stat. 5540 (2022).

[68] S.B. 7 Hearing, supra note 63, at 14.

[69] S.B. 947 (No Robo Bosses Act of 2026), 2025–2026 Leg., Reg. Sess. (Cal. 2026), https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202520260SB947. S.B. 947 is Senator McNerney’s successor to S.B. 7, revised to address the concerns raised in the veto message; it is limited to reliance on an ADS for discipline and termination decisions.

[70] S.B. 7 Hearing, supra note 63, at 17.

[71] Id.

[72] Council Bill 120580, supra note 48; App-Based Workers Speak, supra note 48, at 12, 18.

[73] S.B. 7 Hearing, supra note 63, at 18.

[74] S.B. 7 Hearing, supra note 63, at 18.

[75] Minn. Stat. § 181C.04, subd. 4 (permitting a contracted external reviewer where no uninvolved in-house reviewer is available).

[76] Bartosiak & Modliński, supra note 27, at 602.